modelId
stringlengths 4
112
| sha
stringlengths 40
40
| lastModified
stringlengths 24
24
| tags
list | pipeline_tag
stringclasses 29
values | private
bool 1
class | author
stringlengths 2
38
โ | config
null | id
stringlengths 4
112
| downloads
float64 0
36.8M
โ | likes
float64 0
712
โ | library_name
stringclasses 17
values | __index_level_0__
int64 0
38.5k
| readme
stringlengths 0
186k
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
gauravtripathy/distilbert-base-uncased-finetuned-cola | 4a2b0046dea296de4b95487020a4162972be0190 | 2021-10-12T05:57:36.000Z | [
"pytorch",
"tensorboard",
"distilbert",
"text-classification",
"dataset:glue",
"transformers",
"generated_from_trainer",
"license:apache-2.0",
"model-index"
] | text-classification | false | gauravtripathy | null | gauravtripathy/distilbert-base-uncased-finetuned-cola | 4 | null | transformers | 18,600 | ---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- glue
metrics:
- matthews_correlation
model-index:
- name: distilbert-base-uncased-finetuned-cola
results:
- task:
name: Text Classification
type: text-classification
dataset:
name: glue
type: glue
args: cola
metrics:
- name: Matthews Correlation
type: matthews_correlation
value: 0.5264763891845121
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# distilbert-base-uncased-finetuned-cola
This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the glue dataset.
It achieves the following results on the evaluation set:
- Loss: 0.7550
- Matthews Correlation: 0.5265
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 5
### Training results
| Training Loss | Epoch | Step | Validation Loss | Matthews Correlation |
|:-------------:|:-----:|:----:|:---------------:|:--------------------:|
| 0.5296 | 1.0 | 535 | 0.5144 | 0.4215 |
| 0.3504 | 2.0 | 1070 | 0.4903 | 0.5046 |
| 0.2393 | 3.0 | 1605 | 0.6339 | 0.5058 |
| 0.175 | 4.0 | 2140 | 0.7550 | 0.5265 |
| 0.1259 | 5.0 | 2675 | 0.8688 | 0.5259 |
### Framework versions
- Transformers 4.11.3
- Pytorch 1.9.0+cu111
- Datasets 1.12.1
- Tokenizers 0.10.3
|
gayanin/bart-paraphrase-pubmed | 9dcde1b5f209b41e421a28deac0b6576ed2d3ee3 | 2021-11-06T14:48:18.000Z | [
"pytorch",
"tensorboard",
"bart",
"text2text-generation",
"transformers",
"generated_from_trainer",
"license:apache-2.0",
"model-index",
"autotrain_compatible"
] | text2text-generation | false | gayanin | null | gayanin/bart-paraphrase-pubmed | 4 | null | transformers | 18,601 | ---
license: apache-2.0
tags:
- generated_from_trainer
model-index:
- name: bart-paraphrase-pubmed
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# bart-paraphrase-pubmed
This model is a fine-tuned version of [facebook/bart-base](https://huggingface.co/facebook/bart-base) on an unknown dataset.
It achieves the following results on the evaluation set:
- Loss: 0.6340
- Rouge2 Precision: 0.83
- Rouge2 Recall: 0.6526
- Rouge2 Fmeasure: 0.7144
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 40
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Rouge2 Precision | Rouge2 Recall | Rouge2 Fmeasure |
|:-------------:|:-----:|:-----:|:---------------:|:----------------:|:-------------:|:---------------:|
| 0.6613 | 1.0 | 663 | 0.4750 | 0.8321 | 0.6552 | 0.7167 |
| 0.4993 | 2.0 | 1326 | 0.4404 | 0.8366 | 0.6583 | 0.7203 |
| 0.443 | 3.0 | 1989 | 0.4261 | 0.8319 | 0.6562 | 0.7176 |
| 0.3482 | 4.0 | 2652 | 0.4198 | 0.8348 | 0.6571 | 0.7191 |
| 0.3206 | 5.0 | 3315 | 0.4233 | 0.8344 | 0.656 | 0.7183 |
| 0.294 | 6.0 | 3978 | 0.4334 | 0.835 | 0.657 | 0.719 |
| 0.2404 | 7.0 | 4641 | 0.4437 | 0.8334 | 0.6559 | 0.7178 |
| 0.2228 | 8.0 | 5304 | 0.4438 | 0.8348 | 0.6565 | 0.7187 |
| 0.211 | 9.0 | 5967 | 0.4516 | 0.8329 | 0.6549 | 0.717 |
| 0.1713 | 10.0 | 6630 | 0.4535 | 0.8332 | 0.6547 | 0.7169 |
| 0.1591 | 11.0 | 7293 | 0.4763 | 0.8349 | 0.6561 | 0.7184 |
| 0.1555 | 12.0 | 7956 | 0.4824 | 0.8311 | 0.6534 | 0.7153 |
| 0.1262 | 13.0 | 8619 | 0.4883 | 0.8322 | 0.655 | 0.7167 |
| 0.1164 | 14.0 | 9282 | 0.5025 | 0.8312 | 0.6539 | 0.7158 |
| 0.1108 | 15.0 | 9945 | 0.5149 | 0.8321 | 0.6535 | 0.7157 |
| 0.0926 | 16.0 | 10608 | 0.5340 | 0.8315 | 0.6544 | 0.7159 |
| 0.0856 | 17.0 | 11271 | 0.5322 | 0.8306 | 0.6518 | 0.7142 |
| 0.0785 | 18.0 | 11934 | 0.5346 | 0.8324 | 0.6549 | 0.7167 |
| 0.071 | 19.0 | 12597 | 0.5488 | 0.8311 | 0.652 | 0.714 |
| 0.0635 | 20.0 | 13260 | 0.5624 | 0.8287 | 0.6517 | 0.7132 |
| 0.0608 | 21.0 | 13923 | 0.5612 | 0.8299 | 0.6527 | 0.7141 |
| 0.0531 | 22.0 | 14586 | 0.5764 | 0.8283 | 0.6498 | 0.7119 |
| 0.0486 | 23.0 | 15249 | 0.5832 | 0.8298 | 0.6532 | 0.7148 |
| 0.0465 | 24.0 | 15912 | 0.5866 | 0.83 | 0.6522 | 0.7142 |
| 0.0418 | 25.0 | 16575 | 0.5825 | 0.83 | 0.6523 | 0.7141 |
| 0.0391 | 26.0 | 17238 | 0.5997 | 0.8306 | 0.6545 | 0.716 |
| 0.0376 | 27.0 | 17901 | 0.5894 | 0.8315 | 0.6546 | 0.7164 |
| 0.035 | 28.0 | 18564 | 0.6045 | 0.8306 | 0.6529 | 0.7149 |
| 0.0316 | 29.0 | 19227 | 0.6168 | 0.8311 | 0.6546 | 0.7162 |
| 0.0314 | 30.0 | 19890 | 0.6203 | 0.8311 | 0.6552 | 0.7164 |
| 0.0292 | 31.0 | 20553 | 0.6173 | 0.8315 | 0.6548 | 0.7163 |
| 0.0265 | 32.0 | 21216 | 0.6226 | 0.832 | 0.6548 | 0.7166 |
| 0.0274 | 33.0 | 21879 | 0.6264 | 0.8314 | 0.6538 | 0.7155 |
| 0.0247 | 34.0 | 22542 | 0.6254 | 0.8289 | 0.6515 | 0.7132 |
| 0.0238 | 35.0 | 23205 | 0.6254 | 0.8307 | 0.6519 | 0.7142 |
| 0.0232 | 36.0 | 23868 | 0.6295 | 0.8287 | 0.6515 | 0.7133 |
| 0.0215 | 37.0 | 24531 | 0.6326 | 0.8293 | 0.6523 | 0.7138 |
| 0.0212 | 38.0 | 25194 | 0.6332 | 0.8295 | 0.6522 | 0.714 |
| 0.0221 | 39.0 | 25857 | 0.6335 | 0.8305 | 0.6528 | 0.7147 |
| 0.0202 | 40.0 | 26520 | 0.6340 | 0.83 | 0.6526 | 0.7144 |
### Framework versions
- Transformers 4.12.3
- Pytorch 1.9.0+cu111
- Datasets 1.15.1
- Tokenizers 0.10.3
|
gchhablani/fnet-base-finetuned-mrpc | b9981c8a53c2b87faee181161093af55e6a78b47 | 2021-09-20T09:06:55.000Z | [
"pytorch",
"tensorboard",
"fnet",
"text-classification",
"en",
"dataset:glue",
"arxiv:2105.03824",
"transformers",
"generated_from_trainer",
"fnet-bert-base-comparison",
"license:apache-2.0",
"model-index"
] | text-classification | false | gchhablani | null | gchhablani/fnet-base-finetuned-mrpc | 4 | null | transformers | 18,602 | ---
language:
- en
license: apache-2.0
tags:
- generated_from_trainer
- fnet-bert-base-comparison
datasets:
- glue
metrics:
- accuracy
- f1
model-index:
- name: fnet-base-finetuned-mrpc
results:
- task:
name: Text Classification
type: text-classification
dataset:
name: GLUE MRPC
type: glue
args: mrpc
metrics:
- name: Accuracy
type: accuracy
value: 0.7720588235294118
- name: F1
type: f1
value: 0.8502415458937198
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# fnet-base-finetuned-mrpc
This model is a fine-tuned version of [google/fnet-base](https://huggingface.co/google/fnet-base) on the GLUE MRPC dataset.
It achieves the following results on the evaluation set:
- Loss: 0.9653
- Accuracy: 0.7721
- F1: 0.8502
- Combined Score: 0.8112
The model was fine-tuned to compare [google/fnet-base](https://huggingface.co/google/fnet-base) as introduced in [this paper](https://arxiv.org/abs/2105.03824) against [bert-base-cased](https://huggingface.co/bert-base-cased).
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
This model is trained using the [run_glue](https://github.com/huggingface/transformers/blob/master/examples/pytorch/text-classification/run_glue.py) script. The following command was used:
```bash
#!/usr/bin/bash
python ../run_glue.py \\n --model_name_or_path google/fnet-base \\n --task_name mrpc \\n --do_train \\n --do_eval \\n --max_seq_length 512 \\n --per_device_train_batch_size 16 \\n --learning_rate 2e-5 \\n --num_train_epochs 5 \\n --output_dir fnet-base-finetuned-mrpc \\n --push_to_hub \\n --hub_strategy all_checkpoints \\n --logging_strategy epoch \\n --save_strategy epoch \\n --evaluation_strategy epoch \\n```
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 5.0
### Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 | Combined Score |
|:-------------:|:-----:|:----:|:---------------:|:--------:|:------:|:--------------:|
| 0.544 | 1.0 | 230 | 0.5272 | 0.7328 | 0.8300 | 0.7814 |
| 0.4034 | 2.0 | 460 | 0.6211 | 0.7255 | 0.8298 | 0.7776 |
| 0.2602 | 3.0 | 690 | 0.9110 | 0.7230 | 0.8306 | 0.7768 |
| 0.1688 | 4.0 | 920 | 0.8640 | 0.7696 | 0.8489 | 0.8092 |
| 0.0913 | 5.0 | 1150 | 0.9653 | 0.7721 | 0.8502 | 0.8112 |
### Framework versions
- Transformers 4.11.0.dev0
- Pytorch 1.9.0
- Datasets 1.12.1
- Tokenizers 0.10.3
|
gchhablani/fnet-large-finetuned-cola-copy3 | 4b29da49bc2c6ef6784c60484cdc76c5380429e0 | 2021-10-10T11:08:30.000Z | [
"pytorch",
"tensorboard",
"fnet",
"text-classification",
"en",
"dataset:glue",
"transformers",
"generated_from_trainer",
"license:apache-2.0",
"model-index"
] | text-classification | false | gchhablani | null | gchhablani/fnet-large-finetuned-cola-copy3 | 4 | null | transformers | 18,603 | ---
language:
- en
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- glue
metrics:
- matthews_correlation
model-index:
- name: fnet-large-finetuned-cola-copy3
results:
- task:
name: Text Classification
type: text-classification
dataset:
name: GLUE COLA
type: glue
args: cola
metrics:
- name: Matthews Correlation
type: matthews_correlation
value: 0.0
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# fnet-large-finetuned-cola-copy3
This model is a fine-tuned version of [google/fnet-large](https://huggingface.co/google/fnet-large) on the GLUE COLA dataset.
It achieves the following results on the evaluation set:
- Loss: 0.6554
- Matthews Correlation: 0.0
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 4
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 3.0
### Training results
| Training Loss | Epoch | Step | Validation Loss | Matthews Correlation |
|:-------------:|:-----:|:----:|:---------------:|:--------------------:|
| 0.6408 | 1.0 | 2138 | 0.7329 | 0.0 |
| 0.6589 | 2.0 | 4276 | 0.6311 | 0.0 |
| 0.6467 | 3.0 | 6414 | 0.6554 | 0.0 |
### Framework versions
- Transformers 4.11.0.dev0
- Pytorch 1.9.0
- Datasets 1.12.1
- Tokenizers 0.10.3
|
gchhablani/fnet-large-finetuned-mnli | f7c7766e0095e1bc36a1012c7543820a7c0b1cb6 | 2021-09-24T10:58:04.000Z | [
"pytorch",
"tensorboard",
"fnet",
"text-classification",
"transformers"
] | text-classification | false | gchhablani | null | gchhablani/fnet-large-finetuned-mnli | 4 | null | transformers | 18,604 | Entry not found |
geckos/bert-base-uncased-finetuned-glue-cola | 3b78e79a77a8615cc10927d360533cd437b2c520 | 2021-11-12T11:13:56.000Z | [
"pytorch",
"bert",
"text-classification",
"transformers"
] | text-classification | false | geckos | null | geckos/bert-base-uncased-finetuned-glue-cola | 4 | 1 | transformers | 18,605 | Entry not found |
giacomomiolo/bluebert_reupload | 0815d86f099cfaa2c27465146be55608504a96f4 | 2021-05-19T17:17:05.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"transformers"
] | null | false | giacomomiolo | null | giacomomiolo/bluebert_reupload | 4 | null | transformers | 18,606 | Entry not found |
giacomomiolo/electramed_small_scivocab | d66f0eec09a066c59818c942ff2c2eecaf2718ca | 2020-09-20T14:58:01.000Z | [
"pytorch",
"tf",
"electra",
"pretraining",
"transformers"
] | null | false | giacomomiolo | null | giacomomiolo/electramed_small_scivocab | 4 | null | transformers | 18,607 | Entry not found |
glasses/densenet169 | deeb22d24e6f5586d744d13c89082bf4484f0f1a | 2021-12-01T07:48:55.000Z | [
"pytorch",
"arxiv:1608.06993",
"transformers"
] | null | false | glasses | null | glasses/densenet169 | 4 | null | transformers | 18,608 | # densenet169
Implementation of DenseNet proposed in [Densely Connected Convolutional
Networks](https://arxiv.org/abs/1608.06993)
Create a default models
``` {.sourceCode .}
DenseNet.densenet121()
DenseNet.densenet161()
DenseNet.densenet169()
DenseNet.densenet201()
```
Examples:
``` {.sourceCode .}
# change activation
DenseNet.densenet121(activation = nn.SELU)
# change number of classes (default is 1000 )
DenseNet.densenet121(n_classes=100)
# pass a different block
DenseNet.densenet121(block=...)
# change the initial convolution
model = DenseNet.densenet121()
model.encoder.gate.conv1 = nn.Conv2d(3, 64, kernel_size=3)
# store each feature
x = torch.rand((1, 3, 224, 224))
model = DenseNet.densenet121()
# first call .features, this will activate the forward hooks and tells the model you'll like to get the features
model.encoder.features
model(torch.randn((1,3,224,224)))
# get the features from the encoder
features = model.encoder.features
print([x.shape for x in features])
# [torch.Size([1, 128, 28, 28]), torch.Size([1, 256, 14, 14]), torch.Size([1, 512, 7, 7]), torch.Size([1, 1024, 7, 7])]
```
|
glasses/eca_resnet50t | 31a2f7ec59fbf9ef55c028754fb4ce6b014c582f | 2021-11-30T20:22:09.000Z | [
"pytorch",
"transformers"
] | null | false | glasses | null | glasses/eca_resnet50t | 4 | null | transformers | 18,609 | Entry not found |
glasses/resnet26 | 38114d9e6730ef324ba8fcc109031d706769cf97 | 2021-11-30T20:06:59.000Z | [
"pytorch",
"dataset:imagenet",
"arxiv:1512.03385",
"arxiv:1812.01187",
"transformers",
"image-classification",
"license:apache-2.0"
] | image-classification | false | glasses | null | glasses/resnet26 | 4 | null | transformers | 18,610 | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# resnet26
Implementation of ResNet proposed in [Deep Residual Learning for Image
Recognition](https://arxiv.org/abs/1512.03385)
``` python
ResNet.resnet18()
ResNet.resnet26()
ResNet.resnet34()
ResNet.resnet50()
ResNet.resnet101()
ResNet.resnet152()
ResNet.resnet200()
Variants (d) proposed in `Bag of Tricks for Image Classification with Convolutional Neural Networks <https://arxiv.org/pdf/1812.01187.pdf`_
ResNet.resnet26d()
ResNet.resnet34d()
ResNet.resnet50d()
# You can construct your own one by chaning `stem` and `block`
resnet101d = ResNet.resnet101(stem=ResNetStemC, block=partial(ResNetBottleneckBlock, shortcut=ResNetShorcutD))
```
Examples:
``` python
# change activation
ResNet.resnet18(activation = nn.SELU)
# change number of classes (default is 1000 )
ResNet.resnet18(n_classes=100)
# pass a different block
ResNet.resnet18(block=SENetBasicBlock)
# change the steam
model = ResNet.resnet18(stem=ResNetStemC)
change shortcut
model = ResNet.resnet18(block=partial(ResNetBasicBlock, shortcut=ResNetShorcutD))
# store each feature
x = torch.rand((1, 3, 224, 224))
# get features
model = ResNet.resnet18()
# first call .features, this will activate the forward hooks and tells the model you'll like to get the features
model.encoder.features
model(torch.randn((1,3,224,224)))
# get the features from the encoder
features = model.encoder.features
print([x.shape for x in features])
#[torch.Size([1, 64, 112, 112]), torch.Size([1, 64, 56, 56]), torch.Size([1, 128, 28, 28]), torch.Size([1, 256, 14, 14])]
```
|
glasses/resnet50d | 2acf4629bbe1f8e06444b0bb22a4afc70384dc62 | 2021-11-30T20:10:20.000Z | [
"pytorch",
"dataset:imagenet",
"arxiv:1512.03385",
"arxiv:1812.01187",
"transformers",
"image-classification",
"license:apache-2.0"
] | image-classification | false | glasses | null | glasses/resnet50d | 4 | null | transformers | 18,611 | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# resnet50d
Implementation of ResNet proposed in [Deep Residual Learning for Image
Recognition](https://arxiv.org/abs/1512.03385)
``` python
ResNet.resnet18()
ResNet.resnet26()
ResNet.resnet34()
ResNet.resnet50()
ResNet.resnet101()
ResNet.resnet152()
ResNet.resnet200()
Variants (d) proposed in `Bag of Tricks for Image Classification with Convolutional Neural Networks <https://arxiv.org/pdf/1812.01187.pdf`_
ResNet.resnet26d()
ResNet.resnet34d()
ResNet.resnet50d()
# You can construct your own one by chaning `stem` and `block`
resnet101d = ResNet.resnet101(stem=ResNetStemC, block=partial(ResNetBottleneckBlock, shortcut=ResNetShorcutD))
```
Examples:
``` python
# change activation
ResNet.resnet18(activation = nn.SELU)
# change number of classes (default is 1000 )
ResNet.resnet18(n_classes=100)
# pass a different block
ResNet.resnet18(block=SENetBasicBlock)
# change the steam
model = ResNet.resnet18(stem=ResNetStemC)
change shortcut
model = ResNet.resnet18(block=partial(ResNetBasicBlock, shortcut=ResNetShorcutD))
# store each feature
x = torch.rand((1, 3, 224, 224))
# get features
model = ResNet.resnet18()
# first call .features, this will activate the forward hooks and tells the model you'll like to get the features
model.encoder.features
model(torch.randn((1,3,224,224)))
# get the features from the encoder
features = model.encoder.features
print([x.shape for x in features])
#[torch.Size([1, 64, 112, 112]), torch.Size([1, 64, 56, 56]), torch.Size([1, 128, 28, 28]), torch.Size([1, 256, 14, 14])]
```
|
glasses/resnext50_32x4d | df60bf55e9842db1ab859238f2ff3c141efe07b1 | 2021-11-30T20:13:20.000Z | [
"pytorch",
"arxiv:1611.05431",
"transformers"
] | null | false | glasses | null | glasses/resnext50_32x4d | 4 | null | transformers | 18,612 | # resnext50_32x4d
Implementation of ResNetXt proposed in [\"Aggregated Residual
Transformation for Deep Neural
Networks\"](https://arxiv.org/pdf/1611.05431.pdf)
Create a default model
``` python
ResNetXt.resnext50_32x4d()
ResNetXt.resnext101_32x8d()
# create a resnetxt18_32x4d
ResNetXt.resnet18(block=ResNetXtBottleNeckBlock, groups=32, base_width=4)
```
Examples:
: ``` python
# change activation
ResNetXt.resnext50_32x4d(activation = nn.SELU)
# change number of classes (default is 1000 )
ResNetXt.resnext50_32x4d(n_classes=100)
# pass a different block
ResNetXt.resnext50_32x4d(block=SENetBasicBlock)
# change the initial convolution
model = ResNetXt.resnext50_32x4d
model.encoder.gate.conv1 = nn.Conv2d(3, 64, kernel_size=3)
# store each feature
x = torch.rand((1, 3, 224, 224))
model = ResNetXt.resnext50_32x4d()
# first call .features, this will activate the forward hooks and tells the model you'll like to get the features
model.encoder.features
model(torch.randn((1,3,224,224)))
# get the features from the encoder
features = model.encoder.features
print([x.shape for x in features])
#[torch.Size([1, 64, 112, 112]), torch.Size([1, 64, 56, 56]), torch.Size([1, 128, 28, 28]), torch.Size([1, 256, 14, 14])]
```
|
Language-Media-Lab/mt5-small-jpn-ain-mt | 0533e3cb56fbb394146d9064e1f298bf6a6f5c11 | 2022-02-04T14:23:13.000Z | [
"pytorch",
"mt5",
"text2text-generation",
"jpn",
"ain",
"transformers",
"translation",
"autotrain_compatible"
] | translation | false | Language-Media-Lab | null | Language-Media-Lab/mt5-small-jpn-ain-mt | 4 | null | transformers | 18,613 | ---
language:
- jpn
- ain
tags:
- translation
---
mt5-small-jpn-ain-mt is a machine translation model pretrained with [Google's mT5-small](https://huggingface.co/google/mt5-small) and fine-tuned on bilingual datasets crawled from the Web. It translates Japanese to Ainu language.
|
google/multiberts-seed_4-step_0k | d0cefc34c7342333ad1f2af117db755fc83bfd8c | 2021-11-06T03:01:38.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_0k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_0k | 4 | null | transformers | 18,614 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_0k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 0k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 0k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_0k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_0k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_0k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_0k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_1100k | e0d0c638b01b2c12ec620f5aca946c67b6817bd6 | 2021-11-06T03:35:41.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_1100k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_1100k | 4 | null | transformers | 18,615 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_1100k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 1100k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 1100k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_1100k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_1100k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_1100k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_1100k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_140k | f142a0b740587d8494dcbd60681e788fa7913b14 | 2021-11-06T03:14:03.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_140k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_140k | 4 | null | transformers | 18,616 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_140k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 140k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 140k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_140k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_140k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_140k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_140k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_160k | 3a8c1911c76b30aeb85e06fee3b68c376c0e1b94 | 2021-11-06T03:16:09.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_160k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_160k | 4 | null | transformers | 18,617 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_160k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 160k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 160k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_160k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_160k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_160k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_160k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_1800k | 6b44aad78df72d58bbc1b4bc667718a413ca1320 | 2021-11-06T03:47:28.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_1800k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_1800k | 4 | null | transformers | 18,618 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_1800k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 1800k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 1800k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_1800k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_1800k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_1800k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_1800k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_200k | e0406550e68315dcae7873b3d43768828daf9743 | 2021-11-06T03:19:47.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_200k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_200k | 4 | null | transformers | 18,619 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_200k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 200k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 200k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_200k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_200k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_200k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_200k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_40k | 55480c269dacd05839ab88ca40391f8619a836a6 | 2021-11-06T03:05:04.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_40k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_40k | 4 | null | transformers | 18,620 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_40k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 40k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 40k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_40k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_40k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_40k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_40k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_500k | 48894a2cd8822749f104a399a2bf7c4d2fc2e5c9 | 2021-11-06T03:25:17.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_500k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_500k | 4 | null | transformers | 18,621 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_500k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 500k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 500k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_500k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_500k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_500k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_500k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_80k | 389197ff3ee481eb235cfbfd2fdc17e0374640f3 | 2021-11-06T03:09:10.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_80k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_80k | 4 | null | transformers | 18,622 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_80k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 80k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 80k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_80k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_80k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_80k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_80k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/multiberts-seed_4-step_900k | 3815587336cce1137176d4410b9f350efac609c4 | 2021-11-06T03:32:21.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_4",
"multiberts-seed_4-step_900k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_4-step_900k | 4 | null | transformers | 18,623 | ---
language: en
tags:
- multiberts
- multiberts-seed_4
- multiberts-seed_4-step_900k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 4, Step 900k
MultiBERTs is a collection of checkpoints and a statistical library to support
robust research on BERT. We provide 25 BERT-base models trained with
similar hyper-parameters as
[the original BERT model](https://github.com/google-research/bert) but
with different random seeds, which causes variations in the initial weights and order of
training instances. The aim is to distinguish findings that apply to a specific
artifact (i.e., a particular instance of the model) from those that apply to the
more general procedure.
We also provide 140 intermediate checkpoints captured
during the course of pre-training (we saved 28 checkpoints for the first 5 runs).
The models were originally released through
[http://goo.gle/multiberts](http://goo.gle/multiberts). We describe them in our
paper
[The MultiBERTs: BERT Reproductions for Robustness Analysis](https://arxiv.org/abs/2106.16163).
This is model #4, captured at step 900k (max: 2000k, i.e., 2M steps).
## Model Description
This model was captured during a reproduction of
[BERT-base uncased](https://github.com/google-research/bert), for English: it
is a Transformers model pretrained on a large corpus of English data, using the
Masked Language Modelling (MLM) and the Next Sentence Prediction (NSP)
objectives.
The intended uses, limitations, training data and training procedure for the fully trained model are similar
to [BERT-base uncased](https://github.com/google-research/bert). Two major
differences with the original model:
* We pre-trained the MultiBERTs models for 2 million steps using sequence
length 512 (instead of 1 million steps using sequence length 128 then 512).
* We used an alternative version of Wikipedia and Books Corpus, initially
collected for [Turc et al., 2019](https://arxiv.org/abs/1908.08962).
This is a best-effort reproduction, and so it is probable that differences with
the original model have gone unnoticed. The performance of MultiBERTs on GLUE after full training is oftentimes comparable to that of original
BERT, but we found significant differences on the dev set of SQuAD (MultiBERTs outperforms original BERT).
See our [technical report](https://arxiv.org/abs/2106.16163) for more details.
### How to use
Using code from
[BERT-base uncased](https://huggingface.co/bert-base-uncased), here is an example based on
Tensorflow:
```
from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_900k')
model = TFBertModel.from_pretrained("google/multiberts-seed_4-step_900k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)
```
PyTorch version:
```
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('google/multiberts-seed_4-step_900k')
model = BertModel.from_pretrained("google/multiberts-seed_4-step_900k")
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Citation info
```bibtex
@article{sellam2021multiberts,
title={The MultiBERTs: BERT Reproductions for Robustness Analysis},
author={Thibault Sellam and Steve Yadlowsky and Jason Wei and Naomi Saphra and Alexander D'Amour and Tal Linzen and Jasmijn Bastings and Iulia Turc and Jacob Eisenstein and Dipanjan Das and Ian Tenney and Ellie Pavlick},
journal={arXiv preprint arXiv:2106.16163},
year={2021}
}
```
|
google/t5-efficient-base-dm512 | 02e3acc44d26c7cc80d86114efcfe9beb7b5d81b | 2022-02-15T10:52:25.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-base-dm512 | 4 | null | transformers | 18,624 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-BASE-DM512 (Deep-Narrow version)
T5-Efficient-BASE-DM512 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dm512** - is of model type **Base** with the following variations:
- **dm** is **512**
It has **148.63** million parameters and thus requires *ca.* **594.52 MB** of memory in full precision (*fp32*)
or **297.26 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-base-el6 | 81282706faa482718c784515bda727af717ef73a | 2022-02-15T10:52:34.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-base-el6 | 4 | null | transformers | 18,625 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-BASE-EL6 (Deep-Narrow version)
T5-Efficient-BASE-EL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-el6** - is of model type **Base** with the following variations:
- **el** is **6**
It has **180.45** million parameters and thus requires *ca.* **721.8 MB** of memory in full precision (*fp32*)
or **360.9 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-base-kv16 | a0ec2b0559409d98ebe9b43145708b41e3a2dc52 | 2022-02-15T10:52:59.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-base-kv16 | 4 | null | transformers | 18,626 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-BASE-KV16 (Deep-Narrow version)
T5-Efficient-BASE-KV16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-kv16** - is of model type **Base** with the following variations:
- **kv** is **16**
It has **159.23** million parameters and thus requires *ca.* **636.92 MB** of memory in full precision (*fp32*)
or **318.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-base-kv32 | 430da3dae660175ded17d2250ec8afef2ef3a6b2 | 2022-02-15T10:53:06.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-base-kv32 | 4 | null | transformers | 18,627 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-BASE-KV32 (Deep-Narrow version)
T5-Efficient-BASE-KV32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-kv32** - is of model type **Base** with the following variations:
- **kv** is **32**
It has **180.46** million parameters and thus requires *ca.* **721.86 MB** of memory in full precision (*fp32*)
or **360.93 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-base-nh24 | 07580d6b69521474992b013198c4c7da910cc02d | 2022-02-15T10:53:12.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-base-nh24 | 4 | null | transformers | 18,628 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-BASE-NH24 (Deep-Narrow version)
T5-Efficient-BASE-NH24 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nh24** - is of model type **Base** with the following variations:
- **nh** is **24**
It has **307.87** million parameters and thus requires *ca.* **1231.47 MB** of memory in full precision (*fp32*)
or **615.73 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-base-nh32 | 28f0fcfdced08d73641a5fff722b231884ef0246 | 2022-02-15T10:53:15.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-base-nh32 | 4 | null | transformers | 18,629 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-BASE-NH32 (Deep-Narrow version)
T5-Efficient-BASE-NH32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nh32** - is of model type **Base** with the following variations:
- **nh** is **32**
It has **364.49** million parameters and thus requires *ca.* **1457.96 MB** of memory in full precision (*fp32*)
or **728.98 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-large-dl8 | 1c257b4646b612341dbd091b32dcec0b57ba46f9 | 2022-02-15T10:55:01.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-large-dl8 | 4 | null | transformers | 18,630 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-DL8 (Deep-Narrow version)
T5-Efficient-LARGE-DL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl8** - is of model type **Large** with the following variations:
- **dl** is **8**
It has **469.22** million parameters and thus requires *ca.* **1876.87 MB** of memory in full precision (*fp32*)
or **938.43 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-large-el2 | fae0b2ca961a1292700b141720a2ea81ed94ab2e | 2022-02-15T10:55:07.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-large-el2 | 4 | null | transformers | 18,631 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-EL2 (Deep-Narrow version)
T5-Efficient-LARGE-EL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-el2** - is of model type **Large** with the following variations:
- **el** is **2**
It has **460.84** million parameters and thus requires *ca.* **1843.34 MB** of memory in full precision (*fp32*)
or **921.67 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-large-el6 | ecd0455b7c7d9c4e5b269dcf783995152cb1cbad | 2022-02-15T10:53:52.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-large-el6 | 4 | null | transformers | 18,632 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-EL6 (Deep-Narrow version)
T5-Efficient-LARGE-EL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-el6** - is of model type **Large** with the following variations:
- **el** is **6**
It has **511.18** million parameters and thus requires *ca.* **2044.72 MB** of memory in full precision (*fp32*)
or **1022.36 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-large-kv32 | 99ca70da6abd4c344b218d52f0a2a7509e7df0ab | 2022-02-15T10:55:22.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-large-kv32 | 4 | null | transformers | 18,633 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-KV32 (Deep-Narrow version)
T5-Efficient-LARGE-KV32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-kv32** - is of model type **Large** with the following variations:
- **kv** is **32**
It has **586.73** million parameters and thus requires *ca.* **2346.92 MB** of memory in full precision (*fp32*)
or **1173.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-large-nl12 | 6be18cd8c34453339ae428812f87c0a265945cb6 | 2022-02-15T10:55:44.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-large-nl12 | 4 | null | transformers | 18,634 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-NL12 (Deep-Narrow version)
T5-Efficient-LARGE-NL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl12** - is of model type **Large** with the following variations:
- **nl** is **12**
It has **385.31** million parameters and thus requires *ca.* **1541.25 MB** of memory in full precision (*fp32*)
or **770.63 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-large-nl16 | 4db8b5b975a27441d5c01f15f51d539a33b91e15 | 2022-02-15T10:55:48.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-large-nl16 | 4 | null | transformers | 18,635 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-NL16 (Deep-Narrow version)
T5-Efficient-LARGE-NL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl16** - is of model type **Large** with the following variations:
- **nl** is **16**
It has **502.78** million parameters and thus requires *ca.* **2011.13 MB** of memory in full precision (*fp32*)
or **1005.57 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-mini-nl6 | 20dd96052c4c21f1024f342e888305c2169a6ad7 | 2022-02-15T10:56:14.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-mini-nl6 | 4 | null | transformers | 18,636 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-MINI-NL6 (Deep-Narrow version)
T5-Efficient-MINI-NL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-mini-nl6** - is of model type **Mini** with the following variations:
- **nl** is **6**
It has **40.68** million parameters and thus requires *ca.* **162.7 MB** of memory in full precision (*fp32*)
or **81.35 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-mini-nl8 | e5712b1dd62a78747e7e8a19a07458fc12496bb4 | 2022-02-15T10:56:17.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-mini-nl8 | 4 | null | transformers | 18,637 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-MINI-NL8 (Deep-Narrow version)
T5-Efficient-MINI-NL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-mini-nl8** - is of model type **Mini** with the following variations:
- **nl** is **8**
It has **50.12** million parameters and thus requires *ca.* **200.49 MB** of memory in full precision (*fp32*)
or **100.24 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-dl12 | ab4cb5a1eec7532cbfbd47cf854955fad55c3c92 | 2022-02-15T10:56:23.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-dl12 | 4 | null | transformers | 18,638 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-DL12 (Deep-Narrow version)
T5-Efficient-SMALL-DL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dl12** - is of model type **Small** with the following variations:
- **dl** is **12**
It has **85.7** million parameters and thus requires *ca.* **342.82 MB** of memory in full precision (*fp32*)
or **171.41 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-dm1000 | 3217f0016399d671ee23423fbf8ee706cb2c4362 | 2022-02-15T10:56:40.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-dm1000 | 4 | null | transformers | 18,639 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-DM1000 (Deep-Narrow version)
T5-Efficient-SMALL-DM1000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dm1000** - is of model type **Small** with the following variations:
- **dm** is **1000**
It has **121.03** million parameters and thus requires *ca.* **484.11 MB** of memory in full precision (*fp32*)
or **242.05 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-dm2000 | 0ad0e8c68f6a1f7b7a00da343cd8c641adfe3906 | 2022-02-15T10:50:05.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-dm2000 | 4 | null | transformers | 18,640 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-DM2000 (Deep-Narrow version)
T5-Efficient-SMALL-DM2000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dm2000** - is of model type **Small** with the following variations:
- **dm** is **2000**
It has **242.04** million parameters and thus requires *ca.* **968.16 MB** of memory in full precision (*fp32*)
or **484.08 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-el12 | 25db1d0ba21182973283e12ac15e421e8cb1e9a6 | 2022-02-15T10:56:49.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-el12 | 4 | null | transformers | 18,641 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-EL12 (Deep-Narrow version)
T5-Efficient-SMALL-EL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el12** - is of model type **Small** with the following variations:
- **el** is **12**
It has **79.41** million parameters and thus requires *ca.* **317.63 MB** of memory in full precision (*fp32*)
or **158.81 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-el64 | a183089e246c16288d0d64e87a8720e5e19d4c91 | 2022-02-15T10:54:08.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-el64 | 4 | null | transformers | 18,642 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-EL64 (Deep-Narrow version)
T5-Efficient-SMALL-EL64 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el64** - is of model type **Small** with the following variations:
- **el** is **64**
It has **243.08** million parameters and thus requires *ca.* **972.34 MB** of memory in full precision (*fp32*)
or **486.17 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-kv16 | 99c14571171e83b7ae3ddf92e7dc41eebf8db00b | 2022-02-15T10:50:34.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-kv16 | 4 | null | transformers | 18,643 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-KV16 (Deep-Narrow version)
T5-Efficient-SMALL-KV16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-kv16** - is of model type **Small** with the following variations:
- **kv** is **16**
It has **46.37** million parameters and thus requires *ca.* **185.46 MB** of memory in full precision (*fp32*)
or **92.73 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-kv256 | c3c5308e150d6af87af5aa285816d5624ff8b565 | 2022-02-15T10:57:06.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-kv256 | 4 | null | transformers | 18,644 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-KV256 (Deep-Narrow version)
T5-Efficient-SMALL-KV256 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-kv256** - is of model type **Small** with the following variations:
- **kv** is **256**
It has **117.14** million parameters and thus requires *ca.* **468.58 MB** of memory in full precision (*fp32*)
or **234.29 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-nl16 | e92c36f6cf36d80804bb3ef487df9ce70434c217 | 2022-02-15T10:50:40.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-nl16 | 4 | null | transformers | 18,645 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-NL16 (Deep-Narrow version)
T5-Efficient-SMALL-NL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-nl16** - is of model type **Small** with the following variations:
- **nl** is **16**
It has **133.97** million parameters and thus requires *ca.* **535.88 MB** of memory in full precision (*fp32*)
or **267.94 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-nl2 | 4545fe18983377445abd109eec4e7d1f5b8a69c6 | 2022-02-15T10:50:43.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-nl2 | 4 | null | transformers | 18,646 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-NL2 (Deep-Narrow version)
T5-Efficient-SMALL-NL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-nl2** - is of model type **Small** with the following variations:
- **nl** is **2**
It has **31.14** million parameters and thus requires *ca.* **124.57 MB** of memory in full precision (*fp32*)
or **62.28 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-small-nl40 | c91082cb17de27eeb5cc9a01543f5a8cfc3590b0 | 2022-02-15T10:50:55.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-small-nl40 | 4 | null | transformers | 18,647 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-NL40 (Deep-Narrow version)
T5-Efficient-SMALL-NL40 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-nl40** - is of model type **Small** with the following variations:
- **nl** is **40**
It has **310.25** million parameters and thus requires *ca.* **1240.99 MB** of memory in full precision (*fp32*)
or **620.5 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-tiny-nl12 | f5e456c11ed7976904c611918c2ffce2a81001b4 | 2022-02-15T10:51:34.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-tiny-nl12 | 4 | null | transformers | 18,648 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-TINY-NL12 (Deep-Narrow version)
T5-Efficient-TINY-NL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-tiny-nl12** - is of model type **Tiny** with the following variations:
- **nl** is **12**
It has **30.29** million parameters and thus requires *ca.* **121.16 MB** of memory in full precision (*fp32*)
or **60.58 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-tiny-nl16 | 82d9efbbc5105f8ac4b01ec44ec4534fcd177627 | 2022-02-15T10:49:37.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-tiny-nl16 | 4 | null | transformers | 18,649 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-TINY-NL16 (Deep-Narrow version)
T5-Efficient-TINY-NL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-tiny-nl16** - is of model type **Tiny** with the following variations:
- **nl** is **16**
It has **37.64** million parameters and thus requires *ca.* **150.57 MB** of memory in full precision (*fp32*)
or **75.29 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-tiny-nl2 | dd7593ed67e0599b3774889150e3bb5a27f8324d | 2022-02-15T10:51:37.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-tiny-nl2 | 4 | null | transformers | 18,650 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-TINY-NL2 (Deep-Narrow version)
T5-Efficient-TINY-NL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-tiny-nl2** - is of model type **Tiny** with the following variations:
- **nl** is **2**
It has **11.9** million parameters and thus requires *ca.* **47.61 MB** of memory in full precision (*fp32*)
or **23.81 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-xl-nl2 | 463cd5d9a07b28f633d7d9da935fcf963baa6de3 | 2022-02-15T10:51:53.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-xl-nl2 | 4 | null | transformers | 18,651 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-XL-NL2 (Deep-Narrow version)
T5-Efficient-XL-NL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-xl-nl2** - is of model type **Xl** with the following variations:
- **nl** is **2**
It has **267.8** million parameters and thus requires *ca.* **1071.2 MB** of memory in full precision (*fp32*)
or **535.6 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/t5-efficient-xl-nl4 | ca4325da9002ac1ae351189bc0e6a346bd14d981 | 2022-02-15T10:49:43.000Z | [
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"arxiv:2109.10686",
"transformers",
"deep-narrow",
"license:apache-2.0",
"autotrain_compatible"
] | text2text-generation | false | google | null | google/t5-efficient-xl-nl4 | 4 | null | transformers | 18,652 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-XL-NL4 (Deep-Narrow version)
T5-Efficient-XL-NL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the modelโs depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-xl-nl4** - is of model type **Xl** with the following variations:
- **nl** is **4**
It has **502.7** million parameters and thus requires *ca.* **2010.78 MB** of memory in full precision (*fp32*)
or **1005.39 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future. |
google/tapas-base-masklm | 9609c28345f4d16232489d87582b4e0123d2e99c | 2021-11-29T14:25:30.000Z | [
"pytorch",
"tf",
"tapas",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | google | null | google/tapas-base-masklm | 4 | null | transformers | 18,653 | This model corresponds to **tapas_masklm_base_reset** of the [original repository](https://github.com/google-research/tapas).
Here's how you can use it:
```python
from transformers import TapasTokenizer, TapasForMaskedLM
import pandas as pd
import torch
tokenizer = TapasTokenizer.from_pretrained("google/tapas-base-masklm")
model = TapasForMaskedLM.from_pretrained("google/tapas-base-masklm")
data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"],
'Age': ["56", "45", "59"],
'Number of movies': ["87", "53", "69"]
}
table = pd.DataFrame.from_dict(data)
query = "How many movies has Leonardo [MASK] Caprio played in?"
# prepare inputs
inputs = tokenizer(table=table, queries=query, padding="max_length", return_tensors="pt")
# forward pass
outputs = model(**inputs)
# return top 5 values and predictions
masked_index = torch.nonzero(inputs.input_ids.squeeze() == tokenizer.mask_token_id, as_tuple=False)
logits = outputs.logits[0, masked_index.item(), :]
probs = logits.softmax(dim=0)
values, predictions = probs.topk(5)
for value, pred in zip(values, predictions):
print(f"{tokenizer.decode([pred])} with confidence {value}")
``` |
gorkemgoknar/wav2vec2-large-xlsr-53-turkish | e7072fb151c27ef6aad9c16fe016d14a09070039 | 2021-07-06T05:41:36.000Z | [
"pytorch",
"jax",
"wav2vec2",
"automatic-speech-recognition",
"tr",
"dataset:common_voice",
"dataset:movies",
"transformers",
"audio",
"speech",
"xlsr-fine-tuning-week",
"license:apache-2.0",
"model-index"
] | automatic-speech-recognition | false | gorkemgoknar | null | gorkemgoknar/wav2vec2-large-xlsr-53-turkish | 4 | 1 | transformers | 18,654 | ---
language:
- tr
datasets:
- common_voice
- movies
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Large Turkish with extended dataset by Gorkem Goknar
results:
- task:
name: Speech Recognition
type: automatic-speech-recognition
dataset:
name: Common Voice tr
type: common_voice
args: tr
metrics:
- name: Test WER
type: wer
value: 50.41
---
# Wav2Vec2-Large-XLSR-53-Turkish
Note: This model is trained with 5 Turkish movies additional to common voice dataset.
Although WER is high (50%) per common voice test dataset, performance from "other sources " seems pretty good.
Disclaimer: Please use another wav2vec2-tr model in hub for "clean environment" dialogues as they tend to do better in clean sounds with less background noise.
Dataset building from csv and merging code can be found on below of this Readme.
Please try speech yourself on the right side to see its performance.
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Turkish using the [Common Voice](https://huggingface.co/datasets/common_voice) and 5 Turkish movies that include background noise/talkers .
When using this model, make sure that your speech input is sampled at 16kHz.
## Usage
The model can be used directly (without a language model) as follows:
```python
import torch
import torchaudio
import pydub
from pydub.utils import mediainfo
import array
from pydub import AudioSegment
from pydub.utils import get_array_type
import numpy as np
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
test_dataset = load_dataset("common_voice", "tr", split="test[:2%]")
processor = Wav2Vec2Processor.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish")
model = Wav2Vec2ForCTC.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish")
new_sample_rate = 16000
def audio_resampler(batch, new_sample_rate = 16000):
#not working without complex library compilation in windows for mp3
#speech_array, sampling_rate = torchaudio.load(batch["path"])
#speech_array, sampling_rate = librosa.load(batch["path"])
#sampling_rate = pydub.utils.info['sample_rate'] ##gets current samplerate
sound = pydub.AudioSegment.from_file(file=batch["path"])
sampling_rate = new_sample_rate
sound = sound.set_frame_rate(new_sample_rate)
left = sound.split_to_mono()[0]
bit_depth = left.sample_width * 8
array_type = pydub.utils.get_array_type(bit_depth)
numeric_array = np.array(array.array(array_type, left._data) )
speech_array = torch.FloatTensor(numeric_array)
batch["speech"] = numeric_array
batch["sampling_rate"] = sampling_rate
#batch["target_text"] = batch["sentence"]
return batch
# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
batch = audio_resampler(batch, new_sample_rate = new_sample_rate)
return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])
```
## Evaluation
The model can be evaluated as follows on the Turkish test data of Common Voice.
```python
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
import pydub
import array
import numpy as np
test_dataset = load_dataset("common_voice", "tr", split="test")
wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish")
model = Wav2Vec2ForCTC.from_pretrained("gorkemgoknar/wav2vec2-large-xlsr-53-turkish")
model.to("cuda")
#Note: Not ignoring "'" on this one
#Note: Not ignoring "'" on this one
chars_to_ignore_regex = '[\\,\\?\\.\\!\\-\\;\\:\\"\\โ\\%\\โ\\โ\\๏ฟฝ\\#\\>\\<\\_\\โ\\[\\]\\{\\}]'
#resampler = torchaudio.transforms.Resample(48_000, 16_000)
#using custom load and transformer for audio -> see audio_resampler
new_sample_rate = 16000
def audio_resampler(batch, new_sample_rate = 16000):
#not working without complex library compilation in windows for mp3
#speech_array, sampling_rate = torchaudio.load(batch["path"])
#speech_array, sampling_rate = librosa.load(batch["path"])
#sampling_rate = pydub.utils.info['sample_rate'] ##gets current samplerate
sound = pydub.AudioSegment.from_file(file=batch["path"])
sound = sound.set_frame_rate(new_sample_rate)
left = sound.split_to_mono()[0]
bit_depth = left.sample_width * 8
array_type = pydub.utils.get_array_type(bit_depth)
numeric_array = np.array(array.array(array_type, left._data) )
speech_array = torch.FloatTensor(numeric_array)
return speech_array, new_sample_rate
def remove_special_characters(batch):
##this one comes from subtitles if additional timestamps not processed -> 00:01:01 00:01:01,33
batch["sentence"] = re.sub('\\b\\d{2}:\\d{2}:\\d{2}(,+\\d{2})?\\b', ' ', batch["sentence"])
##remove all caps in text [AรIKLAMA] etc, do it before..
batch["sentence"] = re.sub('\\[(\\b[A-Z]+\\])', '', batch["sentence"])
##replace three dots (that are inside string with single)
batch["sentence"] = re.sub("([a-zA-Z]+)\\.\\.\\.", r"\\1.", batch["sentence"])
#standart ignore list
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + " "
return batch
# Preprocessing the datasets.
# We need to read the aduio files as arrays
new_sample_rate = 16000
def speech_file_to_array_fn(batch):
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
##speech_array, sampling_rate = torchaudio.load(batch["path"])
##load and conversion done in resampler , takes and returns batch
speech_array, sampling_rate = audio_resampler(batch, new_sample_rate = new_sample_rate)
batch["speech"] = speech_array
batch["sampling_rate"] = sampling_rate
batch["target_text"] = batch["sentence"]
return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)
# Preprocessing the datasets.
# We need to read the aduio files as arrays
def evaluate(batch):
inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["pred_strings"] = processor.batch_decode(pred_ids)
return batch
print("EVALUATING:")
##for 8GB RAM on GPU best is batch_size 2 for windows, 4 may fit in linux only
result = test_dataset.map(evaluate, batched=True, batch_size=2)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
```
**Test Result**: 50.41 %
## Training
The Common Voice `train` and `validation` datasets were used for training. Additional 5 Turkish movies with subtitles also used for training.
Similar training model used as base fine-tuning, additional audio resampler is on above code.
Putting model building and merging code below for reference
```python
import pandas as pd
from datasets import load_dataset, load_metric
import os
from pathlib import Path
from datasets import Dataset
import csv
#Walk all subdirectories of base_set_path and find csv files
base_set_path = r'C:\\dataset_extracts'
csv_files = []
for path, subdirs, files in os.walk(base_set_path):
for name in files:
if name.endswith(".csv"):
deckfile= os.path.join(path, name)
csv_files.append(deckfile)
def get_dataset_from_csv_file(csvfilename,names=['sentence', 'path']):
path = Path(csvfilename)
csv_delimiter="\\t" ##tab seperated, change if something else
##Pandas has bug reading non-ascii file names, make sure use open with encoding
df=pd.read_csv(open(path, 'r', encoding='utf-8'), delimiter=csv_delimiter,header=None , names=names, encoding='utf8')
return Dataset.from_pandas(df)
custom_datasets= []
for csv_file in csv_files:
this_dataset=get_dataset_from_csv_file(csv_file)
custom_datasets.append(this_dataset)
from datasets import concatenate_datasets, load_dataset
from datasets import load_from_disk
# Merge datasets together (from csv files)
dataset_file_path = ".\\dataset_file"
custom_datasets_concat = concatenate_datasets( [dset for dset in custom_datasets] )
#save this one to disk
custom_datasets_concat.save_to_disk( dataset_file_path )
#load back from disk
custom_datasets_from_disk = load_from_disk(dataset_file_path)
```
|
groovychoons/biasmodel | 2f3e5b16342af2fac8ec809f5d7ef7388064730d | 2021-10-24T17:43:35.000Z | [
"pytorch",
"distilbert",
"text-classification",
"transformers"
] | text-classification | false | groovychoons | null | groovychoons/biasmodel | 4 | null | transformers | 18,655 | Entry not found |
grumpy/orcas_roberta_mlm | 945c168750b368c235a2e624bdd3b6e256190d78 | 2021-05-20T16:36:08.000Z | [
"pytorch",
"jax",
"roberta",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | grumpy | null | grumpy/orcas_roberta_mlm | 4 | 1 | transformers | 18,656 | Entry not found |
grumpy/weights | 592a8b500986a91ec47183a41dec524a7d709d66 | 2021-05-20T16:37:59.000Z | [
"pytorch",
"jax",
"roberta",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | grumpy | null | grumpy/weights | 4 | null | transformers | 18,657 | Entry not found |
hackertec/roberta-base-bne-finetuned-amazon_reviews_multi-taller | 954b6186fd37def1ab6ac6248707711e08a038a2 | 2021-08-26T18:26:27.000Z | [
"pytorch",
"tensorboard",
"roberta",
"text-classification",
"dataset:amazon_reviews_multi",
"transformers",
"generated_from_trainer",
"license:cc-by-4.0"
] | text-classification | false | hackertec | null | hackertec/roberta-base-bne-finetuned-amazon_reviews_multi-taller | 4 | null | transformers | 18,658 | ---
license: cc-by-4.0
tags:
- generated_from_trainer
datasets:
- amazon_reviews_multi
metrics:
- accuracy
model_index:
- name: roberta-base-bne-finetuned-amazon_reviews_multi-taller
results:
- task:
name: Text Classification
type: text-classification
dataset:
name: amazon_reviews_multi
type: amazon_reviews_multi
args: es
metric:
name: Accuracy
type: accuracy
value: 0.91125
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# roberta-base-bne-finetuned-amazon_reviews_multi-taller
This model is a fine-tuned version of [BSC-TeMU/roberta-base-bne](https://huggingface.co/BSC-TeMU/roberta-base-bne) on the amazon_reviews_multi dataset.
It achieves the following results on the evaluation set:
- Loss: 0.2463
- Accuracy: 0.9113
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 1
### Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy |
|:-------------:|:-----:|:----:|:---------------:|:--------:|
| 0.2474 | 1.0 | 125 | 0.2463 | 0.9113 |
### Framework versions
- Transformers 4.9.2
- Pytorch 1.9.0+cu102
- Datasets 1.11.0
- Tokenizers 0.10.3
|
hanmaroo/xlm_roberta_large_korquad_v2 | 8ef483c80b156099f5868dea3de03e4c384efd04 | 2021-05-14T03:12:15.000Z | [
"pytorch",
"xlm-roberta",
"feature-extraction",
"transformers"
] | feature-extraction | false | hanmaroo | null | hanmaroo/xlm_roberta_large_korquad_v2 | 4 | null | transformers | 18,659 | Entry not found |
hanseokhyeon/bert-badword-puri-1200-base | 0f21bc17486100e8de9ee8e8728f37adb3ec0a0a | 2021-05-19T18:08:32.000Z | [
"pytorch",
"jax",
"bert",
"text-classification",
"transformers"
] | text-classification | false | hanseokhyeon | null | hanseokhyeon/bert-badword-puri-1200-base | 4 | null | transformers | 18,660 | Entry not found |
harish/EN-AStitchTask1A-BERTBaseCased-FalseTrue-0-3-BEST | e4d608c5328d7c73d4976d1039ff9559a8e58d0c | 2021-09-05T00:15:51.000Z | [
"pytorch",
"bert",
"text-classification",
"transformers"
] | text-classification | false | harish | null | harish/EN-AStitchTask1A-BERTBaseCased-FalseTrue-0-3-BEST | 4 | null | transformers | 18,661 | Entry not found |
harish/EN-AStitchTask1A-RoBERTaBase-FalseTrue-0-0-BEST | 9ef8561c68bf331080442c92dc9025416fdb2bd4 | 2021-09-05T00:24:03.000Z | [
"pytorch",
"roberta",
"text-classification",
"transformers"
] | text-classification | false | harish | null | harish/EN-AStitchTask1A-RoBERTaBase-FalseTrue-0-0-BEST | 4 | null | transformers | 18,662 | Entry not found |
harish/PT-UP-xlmR-FalseFalse-FewShot-2_BEST | 1fd6c85d726410cd8e44833c89f0bc70a78d260d | 2021-08-30T13:50:10.000Z | [
"pytorch",
"xlm-roberta",
"text-classification",
"transformers"
] | text-classification | false | harish | null | harish/PT-UP-xlmR-FalseFalse-FewShot-2_BEST | 4 | null | transformers | 18,663 | Entry not found |
harish/PT-mbert-train-from-test-and-dev-SHORT-FalseTrue-0_2_BEST | e4e98b4b85289195a5b1a61bc3e0aa7f8d47aeff | 2021-05-19T18:40:47.000Z | [
"pytorch",
"jax",
"bert",
"text-classification",
"transformers"
] | text-classification | false | harish | null | harish/PT-mbert-train-from-test-and-dev-SHORT-FalseTrue-0_2_BEST | 4 | null | transformers | 18,664 | Entry not found |
harish/PT-v3-dev-test-all-PreTrain-e10-all | c5baf71ce80bc15e76c2084607eb7cfbe2506831 | 2021-05-19T18:42:52.000Z | [
"pytorch",
"jax",
"bert",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | harish | null | harish/PT-v3-dev-test-all-PreTrain-e10-all | 4 | null | transformers | 18,665 | Entry not found |
harish/PT-v3-dev-test-all-PreTrain-e5-all | 393ef48e5521dc5a702581cf631a840b6a766ec5 | 2021-05-19T18:44:56.000Z | [
"pytorch",
"jax",
"bert",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | harish | null | harish/PT-v3-dev-test-all-PreTrain-e5-all | 4 | null | transformers | 18,666 | Entry not found |
hashk1/EsperBERTo-malgranda | 7cd86a3ae12f8c8ac2666783c51b16c67764449e | 2021-05-20T16:38:46.000Z | [
"pytorch",
"jax",
"roberta",
"fill-mask",
"eo",
"transformers",
"autotrain_compatible"
] | fill-mask | false | hashk1 | null | hashk1/EsperBERTo-malgranda | 4 | null | transformers | 18,667 | ---
language: eo
thumbnail: https://huggingface.co/blog/assets/01_how-to-train/EsperBERTo-thumbnail-v2.png
widget:
- text: "ฤu vi paloras la <mask> Esperanto?"
---
## EsperBERTo: RoBERTa-like Language model trained on Esperanto |
hchc/distilbert-base-uncased-finetuned-cola | e0a70992f671ac29d13ecbc326f69ac26e1f7d23 | 2021-10-28T08:08:19.000Z | [
"pytorch",
"tensorboard",
"distilbert",
"text-classification",
"dataset:glue",
"transformers",
"generated_from_trainer",
"license:apache-2.0",
"model-index"
] | text-classification | false | hchc | null | hchc/distilbert-base-uncased-finetuned-cola | 4 | null | transformers | 18,668 | ---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- glue
metrics:
- matthews_correlation
model-index:
- name: distilbert-base-uncased-finetuned-cola
results:
- task:
name: Text Classification
type: text-classification
dataset:
name: glue
type: glue
args: cola
metrics:
- name: Matthews Correlation
type: matthews_correlation
value: 0.5451837431775948
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# distilbert-base-uncased-finetuned-cola
This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the glue dataset.
It achieves the following results on the evaluation set:
- Loss: 0.8508
- Matthews Correlation: 0.5452
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 5
### Training results
| Training Loss | Epoch | Step | Validation Loss | Matthews Correlation |
|:-------------:|:-----:|:----:|:---------------:|:--------------------:|
| 0.5221 | 1.0 | 535 | 0.5370 | 0.4246 |
| 0.3462 | 2.0 | 1070 | 0.5157 | 0.5183 |
| 0.2332 | 3.0 | 1605 | 0.6324 | 0.5166 |
| 0.1661 | 4.0 | 2140 | 0.7616 | 0.5370 |
| 0.1263 | 5.0 | 2675 | 0.8508 | 0.5452 |
### Framework versions
- Transformers 4.11.3
- Pytorch 1.9.0+cu111
- Datasets 1.14.0
- Tokenizers 0.10.3
|
hcjang1987/distilbert-base-uncased-finetuned-cola | 05e88fd8e861cfe162209d063f00ca6b24630352 | 2021-11-10T07:44:24.000Z | [
"pytorch",
"tensorboard",
"distilbert",
"text-classification",
"dataset:glue",
"transformers",
"generated_from_trainer",
"license:apache-2.0",
"model-index"
] | text-classification | false | hcjang1987 | null | hcjang1987/distilbert-base-uncased-finetuned-cola | 4 | null | transformers | 18,669 | ---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- glue
metrics:
- matthews_correlation
model-index:
- name: distilbert-base-uncased-finetuned-cola
results:
- task:
name: Text Classification
type: text-classification
dataset:
name: glue
type: glue
args: cola
metrics:
- name: Matthews Correlation
type: matthews_correlation
value: 0.5471613867597194
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# distilbert-base-uncased-finetuned-cola
This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the glue dataset.
It achieves the following results on the evaluation set:
- Loss: 0.8657
- Matthews Correlation: 0.5472
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 5
### Training results
| Training Loss | Epoch | Step | Validation Loss | Matthews Correlation |
|:-------------:|:-----:|:----:|:---------------:|:--------------------:|
| 0.527 | 1.0 | 535 | 0.5545 | 0.3893 |
| 0.3518 | 2.0 | 1070 | 0.5170 | 0.4970 |
| 0.2448 | 3.0 | 1605 | 0.6734 | 0.5142 |
| 0.1779 | 4.0 | 2140 | 0.7728 | 0.5466 |
| 0.1339 | 5.0 | 2675 | 0.8657 | 0.5472 |
### Framework versions
- Transformers 4.12.3
- Pytorch 1.9.0+cu111
- Datasets 1.15.1
- Tokenizers 0.10.3
|
hf-internal-testing/tiny-deberta | a09f129192911ebe8ace730a9515b8d4a35acc30 | 2021-07-21T04:56:11.000Z | [
"pytorch",
"deberta",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | hf-internal-testing | null | hf-internal-testing/tiny-deberta | 4 | null | transformers | 18,670 | This is a tiny-deberta random model to be used for basic testing.
|
hf-internal-testing/tiny-random-canine | cba4e6f8f2672eb0ca74c5eef543797a78d14354 | 2021-09-17T19:25:07.000Z | [
"pytorch",
"canine",
"transformers"
] | null | false | hf-internal-testing | null | hf-internal-testing/tiny-random-canine | 4 | null | transformers | 18,671 | Entry not found |
hf-internal-testing/tiny-random-convbert | 7ed073dc63589f6fa7c706c2b04b7961bca62d56 | 2021-07-24T15:14:30.000Z | [
"pytorch",
"tf",
"convbert",
"transformers"
] | null | false | hf-internal-testing | null | hf-internal-testing/tiny-random-convbert | 4 | null | transformers | 18,672 | Entry not found |
hf-internal-testing/tiny-random-luke | 7cb05ba4c62e87f8f17be8527141b627419bd9dc | 2021-09-17T19:24:17.000Z | [
"pytorch",
"luke",
"transformers"
] | null | false | hf-internal-testing | null | hf-internal-testing/tiny-random-luke | 4 | null | transformers | 18,673 | Entry not found |
hf-internal-testing/tiny-random-pegasus | efbd1982e5980a5386607dc82ad4972afe603ee3 | 2021-09-17T19:25:37.000Z | [
"pytorch",
"tf",
"pegasus",
"transformers"
] | null | false | hf-internal-testing | null | hf-internal-testing/tiny-random-pegasus | 4 | null | transformers | 18,674 | Entry not found |
hf-test/xls-r-300m-sv-cv8 | ef742096c8900baa78e96c5db154efb29e42da5d | 2022-03-24T11:55:37.000Z | [
"pytorch",
"tensorboard",
"wav2vec2",
"automatic-speech-recognition",
"sv-SE",
"dataset:mozilla-foundation/common_voice_8_0",
"transformers",
"mozilla-foundation/common_voice_8_0",
"generated_from_trainer",
"robust-speech-event",
"hf-asr-leaderboard",
"license:apache-2.0",
"model-index"
] | automatic-speech-recognition | false | hf-test | null | hf-test/xls-r-300m-sv-cv8 | 4 | 1 | transformers | 18,675 | ---
language:
- sv-SE
license: apache-2.0
tags:
- automatic-speech-recognition
- mozilla-foundation/common_voice_8_0
- generated_from_trainer
- robust-speech-event
- hf-asr-leaderboard
datasets:
- mozilla-foundation/common_voice_8_0
model-index:
- name: XLS-R-300M - Swedish - CV8
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Common Voice 8
type: mozilla-foundation/common_voice_8_0
args: sv-SE
metrics:
- name: Test WER
type: wer
value: 17.1
- name: Test CER
type: cer
value: 5.7
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Robust Speech Event - Dev Data
type: speech-recognition-community-v2/dev_data
args: sv
metrics:
- name: Test WER
type: wer
value: 26.92
- name: Test CER
type: cer
value: 12.53
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
#
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - SV-SE dataset.
It achieves the following results on the evaluation set:
**Without LM**:
- Wer: 0.2465
- Cer: 0.0717
**With LM**:
- Wer: 0.1710
- Cer: 0.0569
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 7.5e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 2000
- num_epochs: 50.0
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|:-------------:|:-----:|:-----:|:---------------:|:------:|
| 3.3224 | 1.37 | 500 | 3.2676 | 1.0 |
| 2.9319 | 2.74 | 1000 | 2.9287 | 1.0000 |
| 2.1173 | 4.11 | 1500 | 1.1478 | 0.8788 |
| 1.6973 | 5.48 | 2000 | 0.6749 | 0.6547 |
| 1.5865 | 6.85 | 2500 | 0.5500 | 0.5634 |
| 1.5094 | 8.22 | 3000 | 0.4840 | 0.5430 |
| 1.4644 | 9.59 | 3500 | 0.4844 | 0.4142 |
| 1.4061 | 10.96 | 4000 | 0.4356 | 0.3808 |
| 1.3584 | 12.33 | 4500 | 0.4192 | 0.3698 |
| 1.3438 | 13.7 | 5000 | 0.3980 | 0.3584 |
| 1.3332 | 15.07 | 5500 | 0.3896 | 0.3572 |
| 1.3025 | 16.44 | 6000 | 0.3835 | 0.3487 |
| 1.2979 | 17.81 | 6500 | 0.3781 | 0.3417 |
| 1.2736 | 19.18 | 7000 | 0.3734 | 0.3270 |
| 1.2415 | 20.55 | 7500 | 0.3637 | 0.3316 |
| 1.2255 | 21.92 | 8000 | 0.3546 | 0.3147 |
| 1.2193 | 23.29 | 8500 | 0.3524 | 0.3196 |
| 1.2104 | 24.66 | 9000 | 0.3403 | 0.3097 |
| 1.1965 | 26.03 | 9500 | 0.3508 | 0.3093 |
| 1.1976 | 27.4 | 10000 | 0.3419 | 0.3071 |
| 1.182 | 28.77 | 10500 | 0.3364 | 0.2963 |
| 1.158 | 30.14 | 11000 | 0.3338 | 0.2932 |
| 1.1414 | 31.51 | 11500 | 0.3376 | 0.2940 |
| 1.1402 | 32.88 | 12000 | 0.3370 | 0.2891 |
| 1.1213 | 34.25 | 12500 | 0.3201 | 0.2874 |
| 1.1207 | 35.62 | 13000 | 0.3261 | 0.2826 |
| 1.1074 | 36.98 | 13500 | 0.3117 | 0.2786 |
| 1.0818 | 38.36 | 14000 | 0.3194 | 0.2776 |
| 1.0889 | 39.73 | 14500 | 0.3188 | 0.2738 |
| 1.0672 | 41.1 | 15000 | 0.3196 | 0.2773 |
| 1.0838 | 42.47 | 15500 | 0.3130 | 0.2739 |
| 1.0553 | 43.83 | 16000 | 0.3165 | 0.2704 |
| 1.0786 | 45.21 | 16500 | 0.3108 | 0.2706 |
| 1.0546 | 46.57 | 17000 | 0.3102 | 0.2677 |
| 1.0425 | 47.94 | 17500 | 0.3115 | 0.2679 |
| 1.0398 | 49.31 | 18000 | 0.3131 | 0.2666 |
### Framework versions
- Transformers 4.16.0.dev0
- Pytorch 1.10.1+cu113
- Datasets 1.18.1.dev0
- Tokenizers 0.10.3
|
hiroshi-matsuda-rit/bert-base-japanese-basic-char-v2 | e9609530cf09ec09ee941de10c5cb6f8e398023e | 2021-09-23T14:49:50.000Z | [
"pytorch",
"bert",
"fill-mask",
"ja",
"dataset:wikipedia",
"transformers",
"license:cc-by-sa-4.0",
"autotrain_compatible"
] | fill-mask | false | hiroshi-matsuda-rit | null | hiroshi-matsuda-rit/bert-base-japanese-basic-char-v2 | 4 | null | transformers | 18,676 | ---
language: ja
license: cc-by-sa-4.0
datasets:
- wikipedia
---
# BERT base Japanese (character-level tokenization with whole word masking, jawiki-20200831)
This pretrained model is almost the same as [cl-tohoku/bert-base-japanese-char-v2](https://huggingface.co/cl-tohoku/bert-base-japanese-char-v2) but do not need `fugashi` or `unidic_lite`.
The only difference is in `word_tokenzer_type` property (specify `basic` instead of `mecab`) in `tokenizer_config.json`.
|
histinct7002/distilbert-base-uncased-finetuned-cola | 6c827f71f10ed00e485e3886f93d226aa32ecfc8 | 2022-02-07T06:18:35.000Z | [
"pytorch",
"tensorboard",
"distilbert",
"text-classification",
"dataset:glue",
"transformers",
"generated_from_trainer",
"license:apache-2.0",
"model-index"
] | text-classification | false | histinct7002 | null | histinct7002/distilbert-base-uncased-finetuned-cola | 4 | null | transformers | 18,677 | ---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- glue
metrics:
- matthews_correlation
model-index:
- name: distilbert-base-uncased-finetuned-cola
results:
- task:
name: Text Classification
type: text-classification
dataset:
name: glue
type: glue
args: cola
metrics:
- name: Matthews Correlation
type: matthews_correlation
value: 0.5290966132843783
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# distilbert-base-uncased-finetuned-cola
This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the glue dataset.
It achieves the following results on the evaluation set:
- Loss: 0.4600
- Matthews Correlation: 0.5291
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 5
### Training results
| Training Loss | Epoch | Step | Validation Loss | Matthews Correlation |
|:-------------:|:-----:|:----:|:---------------:|:--------------------:|
| 0.5227 | 1.0 | 535 | 0.4715 | 0.4678 |
| 0.3493 | 2.0 | 1070 | 0.4600 | 0.5291 |
| 0.2393 | 3.0 | 1605 | 0.6018 | 0.5219 |
| 0.1714 | 4.0 | 2140 | 0.7228 | 0.5245 |
| 0.1289 | 5.0 | 2675 | 0.8154 | 0.5279 |
### Framework versions
- Transformers 4.12.5
- Pytorch 1.5.1
- Datasets 1.18.3
- Tokenizers 0.10.3
|
honguyenminh/old-zhongli | e4f6f9ca3c17d6544dd8a636199972bbd055ebe9 | 2021-09-11T15:14:30.000Z | [
"pytorch",
"gpt2",
"text-generation",
"transformers",
"conversational"
] | conversational | false | honguyenminh | null | honguyenminh/old-zhongli | 4 | null | transformers | 18,678 | ---
tags:
- conversational
---
# Zhongli, but not Zhongli |
howey/bert-base-uncased-qnli | 6677e8852acc7a4b3545c7fc6fdc912101e111a5 | 2021-05-26T08:32:48.000Z | [
"pytorch",
"bert",
"text-classification",
"transformers"
] | text-classification | false | howey | null | howey/bert-base-uncased-qnli | 4 | null | transformers | 18,679 | Entry not found |
huchenxu/timit_phoneme | bf517a482024d4cd0d064023de0f947234790f05 | 2021-12-29T12:51:45.000Z | [
"pytorch",
"tensorboard",
"wav2vec2",
"automatic-speech-recognition",
"transformers"
] | automatic-speech-recognition | false | huchenxu | null | huchenxu/timit_phoneme | 4 | null | transformers | 18,680 | Entry not found |
huggingartists/50-cent | 53f49cf85b41695ca4d524ea8f9a3e936c4d63be | 2022-01-03T17:14:20.000Z | [
"pytorch",
"jax",
"gpt2",
"text-generation",
"en",
"dataset:huggingartists/50-cent",
"transformers",
"huggingartists",
"lyrics",
"lm-head",
"causal-lm"
] | text-generation | false | huggingartists | null | huggingartists/50-cent | 4 | null | transformers | 18,681 | ---
language: en
datasets:
- huggingartists/50-cent
tags:
- huggingartists
- lyrics
- lm-head
- causal-lm
widget:
- text: "I am"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:DISPLAY_1; margin-left: auto; margin-right: auto; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://images.genius.com/2aa85f8fdffe5d0552ff319221fc63e4.959x959x1.jpg')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค HuggingArtists Model ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">50 Cent</div>
<a href="https://genius.com/artists/50-cent">
<div style="text-align: center; font-size: 14px;">@50-cent</div>
</a>
</div>
I was made with [huggingartists](https://github.com/AlekseyKorshuk/huggingartists).
Create your own bot based on your favorite artist with [the demo](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb)!
## How does it work?
To understand how the model was developed, check the [W&B report](https://wandb.ai/huggingartists/huggingartists/reportlist).
## Training data
The model was trained on lyrics from 50 Cent.
Dataset is available [here](https://huggingface.co/datasets/huggingartists/50-cent).
And can be used with:
```python
from datasets import load_dataset
dataset = load_dataset("huggingartists/50-cent")
```
[Explore the data](https://wandb.ai/huggingartists/huggingartists/runs/2iekmvbo/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on 50 Cent's lyrics.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/huggingartists/huggingartists/runs/1zrybuom) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/huggingartists/huggingartists/runs/1zrybuom/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingartists/50-cent')
generator("I am", num_return_sequences=5)
```
Or with Transformers library:
```python
from transformers import AutoTokenizer, AutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained("huggingartists/50-cent")
model = AutoModelWithLMHead.from_pretrained("huggingartists/50-cent")
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Aleksey Korshuk*
[](https://github.com/AlekseyKorshuk)
[](https://twitter.com/intent/follow?screen_name=alekseykorshuk)
[](https://t.me/joinchat/_CQ04KjcJ-4yZTky)
For more details, visit the project repository.
[](https://github.com/AlekseyKorshuk/huggingartists)
|
huggingartists/ddt | 359199d6f3f339f45cfe9acc6c58673ea2802373 | 2021-09-16T14:36:22.000Z | [
"pytorch",
"jax",
"gpt2",
"text-generation",
"en",
"dataset:huggingartists/ddt",
"transformers",
"huggingartists",
"lyrics",
"lm-head",
"causal-lm"
] | text-generation | false | huggingartists | null | huggingartists/ddt | 4 | null | transformers | 18,682 | ---
language: en
datasets:
- huggingartists/ddt
tags:
- huggingartists
- lyrics
- lm-head
- causal-lm
widget:
- text: "I am"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:DISPLAY_1; margin-left: auto; margin-right: auto; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://images.rapgenius.com/avatars/medium/f258b58a22ea31bb81b73395c47e5ba4')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค HuggingArtists Model ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">DDT</div>
<a href="https://genius.com/artists/ddt">
<div style="text-align: center; font-size: 14px;">@ddt</div>
</a>
</div>
I was made with [huggingartists](https://github.com/AlekseyKorshuk/huggingartists).
Create your own bot based on your favorite artist with [the demo](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb)!
## How does it work?
To understand how the model was developed, check the [W&B report](https://wandb.ai/huggingartists/huggingartists/reportlist).
## Training data
The model was trained on lyrics from DDT.
Dataset is available [here](https://huggingface.co/datasets/huggingartists/ddt).
And can be used with:
```python
from datasets import load_dataset
dataset = load_dataset("huggingartists/ddt")
```
[Explore the data](https://wandb.ai/huggingartists/huggingartists/runs/2t9xnx5c/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on DDT's lyrics.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/huggingartists/huggingartists/runs/33zphjtk) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/huggingartists/huggingartists/runs/33zphjtk/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingartists/ddt')
generator("I am", num_return_sequences=5)
```
Or with Transformers library:
```python
from transformers import AutoTokenizer, AutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained("huggingartists/ddt")
model = AutoModelWithLMHead.from_pretrained("huggingartists/ddt")
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Aleksey Korshuk*
[](https://github.com/AlekseyKorshuk)
[](https://twitter.com/intent/follow?screen_name=alekseykorshuk)
[](https://t.me/joinchat/_CQ04KjcJ-4yZTky)
For more details, visit the project repository.
[](https://github.com/AlekseyKorshuk/huggingartists)
|
huggingartists/gunna | 96448fad638a7262774c36e66627e180bc7f5b9c | 2021-09-15T17:15:43.000Z | [
"pytorch",
"jax",
"gpt2",
"text-generation",
"en",
"dataset:huggingartists/gunna",
"transformers",
"huggingartists",
"lyrics",
"lm-head",
"causal-lm"
] | text-generation | false | huggingartists | null | huggingartists/gunna | 4 | null | transformers | 18,683 | ---
language: en
datasets:
- huggingartists/gunna
tags:
- huggingartists
- lyrics
- lm-head
- causal-lm
widget:
- text: "I am"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:DISPLAY_1; margin-left: auto; margin-right: auto; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://images.genius.com/18e3833ac527a4bf14ddf2acef834795.640x640x1.jpg')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค HuggingArtists Model ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Gunna</div>
<a href="https://genius.com/artists/gunna">
<div style="text-align: center; font-size: 14px;">@gunna</div>
</a>
</div>
I was made with [huggingartists](https://github.com/AlekseyKorshuk/huggingartists).
Create your own bot based on your favorite artist with [the demo](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb)!
## How does it work?
To understand how the model was developed, check the [W&B report](https://wandb.ai/huggingartists/huggingartists/reportlist).
## Training data
The model was trained on lyrics from Gunna.
Dataset is available [here](https://huggingface.co/datasets/huggingartists/gunna).
And can be used with:
```python
from datasets import load_dataset
dataset = load_dataset("huggingartists/gunna")
```
[Explore the data](https://wandb.ai/huggingartists/huggingartists/runs/vcyblers/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on Gunna's lyrics.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/huggingartists/huggingartists/runs/3c1xymw6) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/huggingartists/huggingartists/runs/3c1xymw6/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingartists/gunna')
generator("I am", num_return_sequences=5)
```
Or with Transformers library:
```python
from transformers import AutoTokenizer, AutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained("huggingartists/gunna")
model = AutoModelWithLMHead.from_pretrained("huggingartists/gunna")
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Aleksey Korshuk*
[](https://github.com/AlekseyKorshuk)
[](https://twitter.com/intent/follow?screen_name=alekseykorshuk)
[](https://t.me/joinchat/_CQ04KjcJ-4yZTky)
For more details, visit the project repository.
[](https://github.com/AlekseyKorshuk/huggingartists)
|
huggingartists/rihanna | eaef62e0deec1885a6592dd7e15023fa0f32cf73 | 2022-06-21T09:56:57.000Z | [
"pytorch",
"jax",
"gpt2",
"text-generation",
"en",
"dataset:huggingartists/rihanna",
"transformers",
"huggingartists",
"lyrics",
"lm-head",
"causal-lm"
] | text-generation | false | huggingartists | null | huggingartists/rihanna | 4 | null | transformers | 18,684 | ---
language: en
datasets:
- huggingartists/rihanna
tags:
- huggingartists
- lyrics
- lm-head
- causal-lm
widget:
- text: "I am"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:DISPLAY_1; margin-left: auto; margin-right: auto; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://images.genius.com/f83548d76e427d0a4fdcafdf2f62b647.1000x1000x1.png')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค HuggingArtists Model ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Rihanna</div>
<a href="https://genius.com/artists/rihanna">
<div style="text-align: center; font-size: 14px;">@rihanna</div>
</a>
</div>
I was made with [huggingartists](https://github.com/AlekseyKorshuk/huggingartists).
Create your own bot based on your favorite artist with [the demo](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb)!
## How does it work?
To understand how the model was developed, check the [W&B report](https://wandb.ai/huggingartists/huggingartists/reportlist).
## Training data
The model was trained on lyrics from Rihanna.
Dataset is available [here](https://huggingface.co/datasets/huggingartists/rihanna).
And can be used with:
```python
from datasets import load_dataset
dataset = load_dataset("huggingartists/rihanna")
```
[Explore the data](https://wandb.ai/huggingartists/huggingartists/runs/ee6eogks/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on Rihanna's lyrics.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/huggingartists/huggingartists/runs/1mvns7x8) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/huggingartists/huggingartists/runs/1mvns7x8/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingartists/rihanna')
generator("I am", num_return_sequences=5)
```
Or with Transformers library:
```python
from transformers import AutoTokenizer, AutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained("huggingartists/rihanna")
model = AutoModelWithLMHead.from_pretrained("huggingartists/rihanna")
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Aleksey Korshuk*
[](https://github.com/AlekseyKorshuk)
[](https://twitter.com/intent/follow?screen_name=alekseykorshuk)
[](https://t.me/joinchat/_CQ04KjcJ-4yZTky)
For more details, visit the project repository.
[](https://github.com/AlekseyKorshuk/huggingartists)
|
huggingartists/slava-kpss | f82b15e618a656efcf11b05590e2b5853c42a3ab | 2021-09-10T16:22:01.000Z | [
"pytorch",
"jax",
"gpt2",
"text-generation",
"en",
"dataset:huggingartists/slava-kpss",
"transformers",
"huggingartists",
"lyrics",
"lm-head",
"causal-lm"
] | text-generation | false | huggingartists | null | huggingartists/slava-kpss | 4 | null | transformers | 18,685 | ---
language: en
datasets:
- huggingartists/slava-kpss
tags:
- huggingartists
- lyrics
- lm-head
- causal-lm
widget:
- text: "I am"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:DISPLAY_1; margin-left: auto; margin-right: auto; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://images.genius.com/e63e3a804916ed71bf2941ac4e190063.847x847x1.jpg')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค HuggingArtists Model ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">ะกะปะฐะฒะฐ ะะะกะก (Slava KPSS)</div>
<a href="https://genius.com/artists/slava-kpss">
<div style="text-align: center; font-size: 14px;">@slava-kpss</div>
</a>
</div>
I was made with [huggingartists](https://github.com/AlekseyKorshuk/huggingartists).
Create your own bot based on your favorite artist with [the demo](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb)!
## How does it work?
To understand how the model was developed, check the [W&B report](https://wandb.ai/huggingartists/huggingartists/reportlist).
## Training data
The model was trained on lyrics from ะกะปะฐะฒะฐ ะะะกะก (Slava KPSS).
Dataset is available [here](https://huggingface.co/datasets/huggingartists/slava-kpss).
And can be used with:
```python
from datasets import load_dataset
dataset = load_dataset("huggingartists/slava-kpss")
```
[Explore the data](https://wandb.ai/huggingartists/huggingartists/runs/2f2r3u3b/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on ะกะปะฐะฒะฐ ะะะกะก (Slava KPSS)'s lyrics.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/huggingartists/huggingartists/runs/pecxkpae) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/huggingartists/huggingartists/runs/pecxkpae/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingartists/slava-kpss')
generator("I am", num_return_sequences=5)
```
Or with Transformers library:
```python
from transformers import AutoTokenizer, AutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained("huggingartists/slava-kpss")
model = AutoModelWithLMHead.from_pretrained("huggingartists/slava-kpss")
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Aleksey Korshuk*
[](https://github.com/AlekseyKorshuk)
[](https://twitter.com/intent/follow?screen_name=alekseykorshuk)
[](https://t.me/joinchat/_CQ04KjcJ-4yZTky)
For more details, visit the project repository.
[](https://github.com/AlekseyKorshuk/huggingartists)
|
huggingface-course/bert-finetuned-squad-accelerate | 7929a34eae1fe3dc7f39768af31b155569d30c46 | 2021-09-26T20:18:30.000Z | [
"pytorch",
"bert",
"question-answering",
"transformers",
"autotrain_compatible"
] | question-answering | false | huggingface-course | null | huggingface-course/bert-finetuned-squad-accelerate | 4 | null | transformers | 18,686 | Entry not found |
huggingface-course/marian-finetuned-kde4-en-to-fr-accelerate | 55601bd141b42ab507078c2af075d3a4149f5062 | 2021-09-28T23:45:46.000Z | [
"pytorch",
"marian",
"text2text-generation",
"transformers",
"autotrain_compatible"
] | text2text-generation | false | huggingface-course | null | huggingface-course/marian-finetuned-kde4-en-to-fr-accelerate | 4 | null | transformers | 18,687 | Entry not found |
huggingface-course/mt5-finetuned-amazon-en-es | 21937578076b0e42b310de3e39e0a63d1d857ee6 | 2021-11-11T12:47:29.000Z | [
"pytorch",
"tensorboard",
"mt5",
"text2text-generation",
"transformers",
"summarization",
"generated_from_trainer",
"license:apache-2.0",
"model-index",
"autotrain_compatible"
] | summarization | false | huggingface-course | null | huggingface-course/mt5-finetuned-amazon-en-es | 4 | null | transformers | 18,688 | ---
license: apache-2.0
tags:
- summarization
- generated_from_trainer
metrics:
- rouge
model-index:
- name: mt5-finetuned-amazon-en-es
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# mt5-finetuned-amazon-en-es
This model is a fine-tuned version of [google/mt5-small](https://huggingface.co/google/mt5-small) on the None dataset.
It achieves the following results on the evaluation set:
- Loss: 3.0285
- Rouge1: 16.9728
- Rouge2: 8.2969
- Rougel: 16.8366
- Rougelsum: 16.851
- Gen Len: 10.1597
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5.6e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 8
### Training results
| Training Loss | Epoch | Step | Validation Loss | Rouge1 | Rouge2 | Rougel | Rougelsum | Gen Len |
|:-------------:|:-----:|:----:|:---------------:|:-------:|:------:|:-------:|:---------:|:-------:|
| 7.1016 | 1.0 | 1209 | 3.3069 | 13.9858 | 5.8437 | 13.6053 | 13.5125 | 8.3782 |
| 3.898 | 2.0 | 2418 | 3.1567 | 16.6706 | 8.6393 | 16.2882 | 16.2249 | 9.7521 |
| 3.5915 | 3.0 | 3627 | 3.0928 | 17.111 | 8.3921 | 16.9139 | 16.7805 | 10.3445 |
| 3.4174 | 4.0 | 4836 | 3.0482 | 16.9728 | 8.3066 | 16.8868 | 16.8485 | 10.3151 |
| 3.3258 | 5.0 | 6045 | 3.0375 | 16.5972 | 8.2621 | 16.3524 | 16.3093 | 10.0672 |
| 3.2427 | 6.0 | 7254 | 3.0232 | 17.3009 | 8.6087 | 17.0782 | 17.0105 | 10.0756 |
| 3.2009 | 7.0 | 8463 | 3.0302 | 16.9284 | 8.6569 | 16.7885 | 16.7784 | 10.2143 |
| 3.1838 | 8.0 | 9672 | 3.0285 | 16.9728 | 8.2969 | 16.8366 | 16.851 | 10.1597 |
### Framework versions
- Transformers 4.12.3
- Pytorch 1.9.1+cu111
- Datasets 1.15.1
- Tokenizers 0.10.3
|
huggingfans/Electra_small_index_epochs3_maxLen512_loss0.19 | 3ef2b14e7887d638c2346cd848cfcbf77bd05145 | 2020-08-19T19:50:33.000Z | [
"pytorch",
"electra",
"text-classification",
"transformers"
] | text-classification | false | huggingfans | null | huggingfans/Electra_small_index_epochs3_maxLen512_loss0.19 | 4 | null | transformers | 18,689 | Entry not found |
huggingtweets/alampaydavis | 9f678507e9c290d931f77dec97c3cb372b0db18b | 2021-07-22T23:20:00.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/alampaydavis | 4 | null | transformers | 18,690 | ---
language: en
thumbnail: https://www.huggingtweets.com/alampaydavis/1626995945354/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1410861223273549825/HwwcW6y2_400x400.jpg')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI BOT ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Matthew</div>
<div style="text-align: center; font-size: 14px;">@alampaydavis</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from Matthew.
| Data | Matthew |
| --- | --- |
| Tweets downloaded | 3219 |
| Retweets | 1067 |
| Short tweets | 228 |
| Tweets kept | 1924 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/1qawxu8m/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @alampaydavis's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/1ub62hd1) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/1ub62hd1/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/alampaydavis')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/aly__dixon-haleyosomething-svpino | e188081886ef56c21542271c20eb737be72cc239 | 2021-09-26T12:49:07.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/aly__dixon-haleyosomething-svpino | 4 | null | transformers | 18,691 | ---
language: en
thumbnail: https://www.huggingtweets.com/aly__dixon-haleyosomething-svpino/1632660543535/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1416541994952937474/yi5cJxnq_400x400.jpg')">
</div>
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1368667185879584770/pKNxJut-_400x400.jpg')">
</div>
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1393327649318076417/cQWDVv-q_400x400.jpg')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI CYBORG ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">haley o'shaughnessy & Santiago & Aly Dixon</div>
<div style="text-align: center; font-size: 14px;">@aly__dixon-haleyosomething-svpino</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from haley o'shaughnessy & Santiago & Aly Dixon.
| Data | haley o'shaughnessy | Santiago | Aly Dixon |
| --- | --- | --- | --- |
| Tweets downloaded | 3241 | 3250 | 3003 |
| Retweets | 430 | 7 | 426 |
| Short tweets | 460 | 316 | 195 |
| Tweets kept | 2351 | 2927 | 2382 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/1mt8xsda/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @aly__dixon-haleyosomething-svpino's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/31g4nsgq) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/31g4nsgq/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/aly__dixon-haleyosomething-svpino')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/amongusgame | ebf1a73c81ec92e6e12264d6449cad78cab31f26 | 2022-01-27T23:46:01.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/amongusgame | 4 | null | transformers | 18,692 | ---
language: en
thumbnail: http://www.huggingtweets.com/amongusgame/1643327156823/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1455270159975796736/PqmjT7Dr_400x400.jpg')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI BOT ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Among Us</div>
<div style="text-align: center; font-size: 14px;">@amongusgame</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from Among Us.
| Data | Among Us |
| --- | --- |
| Tweets downloaded | 3248 |
| Retweets | 75 |
| Short tweets | 1295 |
| Tweets kept | 1878 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/2pyl3gg2/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @amongusgame's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/1vvgbbml) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/1vvgbbml/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/amongusgame')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/corpsecrusader | 9d95e46fda22ea8d0b5f3485fa53075833b669f8 | 2022-04-29T13:57:10.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/corpsecrusader | 4 | null | transformers | 18,693 | ---
language: en
thumbnail: http://www.huggingtweets.com/corpsecrusader/1651240626010/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1515787050334801925/tyxpMmj1_400x400.jpg')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI BOT ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Corpse Crusader ๐ซ๐ซ๐ฎ gamedev hours๐งฑ๐๐จ๐ช</div>
<div style="text-align: center; font-size: 14px;">@corpsecrusader</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from Corpse Crusader ๐ซ๐ซ๐ฎ gamedev hours๐งฑ๐๐จ๐ช.
| Data | Corpse Crusader ๐ซ๐ซ๐ฎ gamedev hours๐งฑ๐๐จ๐ช |
| --- | --- |
| Tweets downloaded | 3244 |
| Retweets | 405 |
| Short tweets | 658 |
| Tweets kept | 2181 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/ogdqtie2/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @corpsecrusader's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/1ecpg08j) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/1ecpg08j/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/corpsecrusader')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/cosm1cgrandma-glitchre-glitchre8 | 632b84c48e3532e32ba4028f253189da67163d59 | 2021-09-14T22:32:11.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/cosm1cgrandma-glitchre-glitchre8 | 4 | null | transformers | 18,694 | ---
language: en
thumbnail: https://www.huggingtweets.com/cosm1cgrandma-glitchre-glitchre8/1631658643977/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1265389720047058944/hWPrCwh7_400x400.jpg')">
</div>
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1394712172010393608/tkWea9AS_400x400.jpg')">
</div>
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1406255548228640781/wzOACSA8_400x400.jpg')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI CYBORG ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">SA | Glitchre & glitchre & cosmic gangster</div>
<div style="text-align: center; font-size: 14px;">@cosm1cgrandma-glitchre-glitchre8</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from SA | Glitchre & glitchre & cosmic gangster.
| Data | SA | Glitchre | glitchre | cosmic gangster |
| --- | --- | --- | --- |
| Tweets downloaded | 2920 | 2891 | 2960 |
| Retweets | 347 | 808 | 1410 |
| Short tweets | 872 | 600 | 359 |
| Tweets kept | 1701 | 1483 | 1191 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/15s2bdg3/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @cosm1cgrandma-glitchre-glitchre8's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/3jv76342) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/3jv76342/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/cosm1cgrandma-glitchre-glitchre8')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/daddykratos1 | 0c2ebf15f5de34dbb3873125733786cc9ca39524 | 2021-08-21T20:04:36.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/daddykratos1 | 4 | null | transformers | 18,695 | ---
language: en
thumbnail: https://www.huggingtweets.com/daddykratos1/1629576272636/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1353401231729950721/EAfnSQDa_400x400.jpg')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI BOT ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Tweets by Kratos๐ช</div>
<div style="text-align: center; font-size: 14px;">@daddykratos1</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from Tweets by Kratos๐ช.
| Data | Tweets by Kratos๐ช |
| --- | --- |
| Tweets downloaded | 626 |
| Retweets | 14 |
| Short tweets | 52 |
| Tweets kept | 560 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/12nz41n2/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @daddykratos1's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/33zt2owy) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/33zt2owy/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/daddykratos1')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/degg-dril-fred_delicious | d946ffa0c37fa184589f628bfe9ecb8f1a4e2ada | 2021-10-21T19:39:06.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/degg-dril-fred_delicious | 4 | null | transformers | 18,696 | ---
language: en
thumbnail: https://www.huggingtweets.com/degg-dril-fred_delicious/1634845142916/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/847818629840228354/VXyQHfn0_400x400.jpg')">
</div>
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/58546628/goat22_400x400.jpg')">
</div>
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/726824334002638848/BEZFr1k8_400x400.jpg')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI CYBORG ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">wint & deg & Fred Delicious</div>
<div style="text-align: center; font-size: 14px;">@degg-dril-fred_delicious</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from wint & deg & Fred Delicious.
| Data | wint | deg | Fred Delicious |
| --- | --- | --- | --- |
| Tweets downloaded | 3227 | 3152 | 3235 |
| Retweets | 473 | 142 | 429 |
| Short tweets | 318 | 42 | 398 |
| Tweets kept | 2436 | 2968 | 2408 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/1mwoed1f/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @degg-dril-fred_delicious's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/1a691ucn) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/1a691ucn/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/degg-dril-fred_delicious')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/factfictyoutube | fb980a796465093411fb18969487213ed070ac9a | 2021-07-29T05:39:17.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/factfictyoutube | 4 | null | transformers | 18,697 | ---
language: en
thumbnail: https://www.huggingtweets.com/factfictyoutube/1627537153556/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1364780488351711234/IEnAKzOo_400x400.jpg')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI BOT ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Factually Fictitious</div>
<div style="text-align: center; font-size: 14px;">@factfictyoutube</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from Factually Fictitious.
| Data | Factually Fictitious |
| --- | --- |
| Tweets downloaded | 3211 |
| Retweets | 957 |
| Short tweets | 134 |
| Tweets kept | 2120 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/b5r6jmml/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @factfictyoutube's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/31b4bo7z) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/31b4bo7z/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/factfictyoutube')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/islamrizza | 81160e6f10f1c9c54c0ffba95baa1db04614bfa9 | 2021-05-22T08:34:49.000Z | [
"pytorch",
"jax",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/islamrizza | 4 | null | transformers | 18,698 | ---
language: en
thumbnail: https://www.huggingtweets.com/islamrizza/1619378181874/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div>
<div style="width: 132px; height:132px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1344394699470082049/YzE4UMsj_400x400.jpg')">
</div>
<div style="margin-top: 8px; font-size: 19px; font-weight: 800">Rizza Islam ๐ค AI Bot </div>
<div style="font-size: 15px">@islamrizza bot</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on [@islamrizza's tweets](https://twitter.com/islamrizza).
| Data | Quantity |
| --- | --- |
| Tweets downloaded | 3195 |
| Retweets | 73 |
| Short tweets | 394 |
| Tweets kept | 2728 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/t09cn5o0/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @islamrizza's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/m6l6wkff) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/m6l6wkff/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/islamrizza')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
huggingtweets/kimkardashian | b9e230dd0fcb18b866b246a4e3c92df09d14f6d8 | 2022-05-29T16:51:46.000Z | [
"pytorch",
"gpt2",
"text-generation",
"en",
"transformers",
"huggingtweets"
] | text-generation | false | huggingtweets | null | huggingtweets/kimkardashian | 4 | null | transformers | 18,699 | ---
language: en
thumbnail: http://www.huggingtweets.com/kimkardashian/1653843073596/predictions.png
tags:
- huggingtweets
widget:
- text: "My dream is"
---
<div class="inline-flex flex-col" style="line-height: 1.5;">
<div class="flex">
<div
style="display:inherit; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('https://pbs.twimg.com/profile_images/1446623190252343301/qIJAwo9I_400x400.jpg')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
<div
style="display:none; margin-left: 4px; margin-right: 4px; width: 92px; height:92px; border-radius: 50%; background-size: cover; background-image: url('')">
</div>
</div>
<div style="text-align: center; margin-top: 3px; font-size: 16px; font-weight: 800">๐ค AI BOT ๐ค</div>
<div style="text-align: center; font-size: 16px; font-weight: 800">Kim Kardashian</div>
<div style="text-align: center; font-size: 14px;">@kimkardashian</div>
</div>
I was made with [huggingtweets](https://github.com/borisdayma/huggingtweets).
Create your own bot based on your favorite user with [the demo](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)!
## How does it work?
The model uses the following pipeline.
To understand how the model was developed, check the [W&B report](https://wandb.ai/wandb/huggingtweets/reports/HuggingTweets-Train-a-Model-to-Generate-Tweets--VmlldzoxMTY5MjI).
## Training data
The model was trained on tweets from Kim Kardashian.
| Data | Kim Kardashian |
| --- | --- |
| Tweets downloaded | 3240 |
| Retweets | 707 |
| Short tweets | 662 |
| Tweets kept | 1871 |
[Explore the data](https://wandb.ai/wandb/huggingtweets/runs/2cfvtzsy/artifacts), which is tracked with [W&B artifacts](https://docs.wandb.com/artifacts) at every step of the pipeline.
## Training procedure
The model is based on a pre-trained [GPT-2](https://huggingface.co/gpt2) which is fine-tuned on @kimkardashian's tweets.
Hyperparameters and metrics are recorded in the [W&B training run](https://wandb.ai/wandb/huggingtweets/runs/3jhtpw08) for full transparency and reproducibility.
At the end of training, [the final model](https://wandb.ai/wandb/huggingtweets/runs/3jhtpw08/artifacts) is logged and versioned.
## How to use
You can use this model directly with a pipeline for text generation:
```python
from transformers import pipeline
generator = pipeline('text-generation',
model='huggingtweets/kimkardashian')
generator("My dream is", num_return_sequences=5)
```
## Limitations and bias
The model suffers from [the same limitations and bias as GPT-2](https://huggingface.co/gpt2#limitations-and-bias).
In addition, the data present in the user's tweets further affects the text generated by the model.
## About
*Built by Boris Dayma*
[](https://twitter.com/intent/follow?screen_name=borisdayma)
For more details, visit the project repository.
[](https://github.com/borisdayma/huggingtweets)
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.