modelId
stringlengths 4
112
| sha
stringlengths 40
40
| lastModified
stringlengths 24
24
| tags
sequence | pipeline_tag
stringclasses 29
values | private
bool 1
class | author
stringlengths 2
38
⌀ | config
null | id
stringlengths 4
112
| downloads
float64 0
36.8M
⌀ | likes
float64 0
712
⌀ | library_name
stringclasses 17
values | __index_level_0__
int64 0
38.5k
| readme
stringlengths 0
186k
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
gchhablani/wav2vec2-large-xlsr-mr | 5be04fd6f46264bd5de20c5e0b466ca687da6097 | 2021-07-06T05:10:15.000Z | [
"pytorch",
"jax",
"wav2vec2",
"automatic-speech-recognition",
"mr",
"dataset:openslr",
"transformers",
"audio",
"speech",
"xlsr-fine-tuning-week",
"license:apache-2.0",
"model-index"
] | automatic-speech-recognition | false | gchhablani | null | gchhablani/wav2vec2-large-xlsr-mr | 1 | null | transformers | 29,000 | ---
language: mr
datasets:
- openslr
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Large 53 Marathi by Gunjan Chhablani
results:
- task:
name: Speech Recognition
type: automatic-speech-recognition
dataset:
name: OpenSLR mr
type: openslr
metrics:
- name: Test WER
type: wer
value: 14.53
---
# Wav2Vec2-Large-XLSR-53-Marathi
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Marathi using the [OpenSLR SLR64](http://openslr.org/64/) dataset. Note that this data contains only female voices. Please keep this in mind before using the model for your task, although it works very well for male voice too. 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, assuming you have a dataset with Marathi `sentence` and `path` fields:
```python
import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
# test_dataset = #TODO: WRITE YOUR CODE TO LOAD THE TEST DATASET. For sample see the Colab link in Training Section.
processor = Wav2Vec2Processor.from_pretrained("gchhablani/wav2vec2-large-xlsr-mr")
model = Wav2Vec2ForCTC.from_pretrained("gchhablani/wav2vec2-large-xlsr-mr")
resampler = torchaudio.transforms.Resample(48_000, 16_000) # The original data was with 48,000 sampling rate. You can change it according to your input.
# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
speech_array, sampling_rate = torchaudio.load(batch["path"])
batch["speech"] = resampler(speech_array).squeeze().numpy()
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 10% of the Marathi data on OpenSLR.
```python
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
# test_dataset = #TODO: WRITE YOUR CODE TO LOAD THE TEST DATASET. For sample see the Colab link in Training Section.
wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("gchhablani/wav2vec2-large-xlsr-mr")
model = Wav2Vec2ForCTC.from_pretrained("gchhablani/wav2vec2-large-xlsr-mr")
model.to("cuda")
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\–\…]'
resampler = torchaudio.transforms.Resample(48_000, 16_000)
# Preprocessing the datasets.
# We need to read the aduio files as arrays
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"])
batch["speech"] = resampler(speech_array).squeeze().numpy()
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
result = test_dataset.map(evaluate, batched=True, batch_size=8)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
```
**Test Result**: 14.53 %
## Training
90% of the OpenSLR Marathi dataset was used for training.
The colab notebook used for training can be found [here](https://colab.research.google.com/drive/1_BbLyLqDUsXG3RpSULfLRjC6UY3RjwME?usp=sharing).
|
gdimino/voxpopuli_base_it_2 | ab2ffe2e0280c807a55f6f79ee37798369c5060d | 2022-05-23T13:46:22.000Z | [
"pytorch",
"wav2vec2",
"automatic-speech-recognition",
"transformers"
] | automatic-speech-recognition | false | gdimino | null | gdimino/voxpopuli_base_it_2 | 1 | null | transformers | 29,001 | Entry not found |
geninhu/roberta_large_ITPT_FP | 5cb29d591609111f9823a6d453b298829f79b6c9 | 2022-02-12T04:48:49.000Z | [
"pytorch",
"roberta",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | geninhu | null | geninhu/roberta_large_ITPT_FP | 1 | null | transformers | 29,002 | Entry not found |
geninhu/xls-asr-vi-40h-1B | 9d78f75c88f418ad886ebcddda2ec922ffe9b173 | 2022-03-23T18:27:57.000Z | [
"pytorch",
"tensorboard",
"wav2vec2",
"automatic-speech-recognition",
"vi",
"dataset:mozilla-foundation/common_voice_7_0",
"transformers",
"common-voice",
"hf-asr-leaderboard",
"robust-speech-event",
"license:apache-2.0",
"model-index"
] | automatic-speech-recognition | false | geninhu | null | geninhu/xls-asr-vi-40h-1B | 1 | null | transformers | 29,003 | ---
license: apache-2.0
language:
- vi
tags:
- automatic-speech-recognition
- common-voice
- hf-asr-leaderboard
- robust-speech-event
datasets:
- mozilla-foundation/common_voice_7_0
model-index:
- name: xls-asr-vi-40h-1B
results:
- task:
name: Speech Recognition
type: automatic-speech-recognition
dataset:
name: Common Voice 7.0
type: mozilla-foundation/common_voice_7_0
args: vi
metrics:
- name: Test WER (with LM)
type: wer
value: 25.846
- name: Test CER (with LM)
type: cer
value: 12.961
- task:
name: Speech Recognition
type: automatic-speech-recognition
dataset:
name: Common Voice 8.0
type: mozilla-foundation/common_voice_8_0
args: vi
metrics:
- name: Test WER (with LM)
type: wer
value: 31.158
- name: Test CER (with LM)
type: cer
value: 16.179
---
# xls-asr-vi-40h-1B
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on 40 hours of FPT Open Speech Dataset (FOSD) and Common Voice 7.0.
### Benchmark WER result:
| | [VIVOS](https://huggingface.co/datasets/vivos) | [COMMON VOICE 7.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) | [COMMON VOICE 8.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_8_0)
|---|---|---|---|
|without LM| 25.93 | 34.21 |
|with 4-grams LM| 24.11 | 25.84 | 31.158 |
### Benchmark CER result:
| | [VIVOS](https://huggingface.co/datasets/vivos) | [COMMON VOICE 7.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) | [COMMON VOICE 8.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_8_0)
|---|---|---|---|
|without LM| 9.24 | 19.94 |
|with 4-grams LM| 10.37 | 12.96 | 16.179 |
## Evaluation
Please use the eval.py file to run the evaluation
```python
python eval.py --model_id geninhu/xls-asr-vi-40h-1B --dataset mozilla-foundation/common_voice_7_0 --config vi --split test --log_outputs
```
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 2
- 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: 1500
- num_epochs: 10.0
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|:-------------:|:-----:|:-----:|:---------------:|:------:|
| 4.6222 | 1.85 | 1500 | 5.9479 | 0.5474 |
| 1.1362 | 3.7 | 3000 | 7.9799 | 0.5094 |
| 0.7814 | 5.56 | 4500 | 5.0330 | 0.4724 |
| 0.6281 | 7.41 | 6000 | 2.3484 | 0.5020 |
| 0.5472 | 9.26 | 7500 | 2.2495 | 0.4793 |
| 0.4827 | 11.11 | 9000 | 1.1530 | 0.4768 |
| 0.4327 | 12.96 | 10500 | 1.6160 | 0.4646 |
| 0.3989 | 14.81 | 12000 | 3.2633 | 0.4703 |
| 0.3522 | 16.67 | 13500 | 2.2337 | 0.4708 |
| 0.3201 | 18.52 | 15000 | 3.6879 | 0.4565 |
| 0.2899 | 20.37 | 16500 | 5.4389 | 0.4599 |
| 0.2776 | 22.22 | 18000 | 3.5284 | 0.4537 |
| 0.2574 | 24.07 | 19500 | 2.1759 | 0.4649 |
| 0.2378 | 25.93 | 21000 | 3.3901 | 0.4448 |
| 0.217 | 27.78 | 22500 | 1.1632 | 0.4565 |
| 0.2115 | 29.63 | 24000 | 1.7441 | 0.4232 |
| 0.1959 | 31.48 | 25500 | 3.4992 | 0.4304 |
| 0.187 | 33.33 | 27000 | 3.6163 | 0.4369 |
| 0.1748 | 35.19 | 28500 | 3.6038 | 0.4467 |
| 0.17 | 37.04 | 30000 | 2.9708 | 0.4362 |
| 0.159 | 38.89 | 31500 | 3.2045 | 0.4279 |
| 0.153 | 40.74 | 33000 | 3.2427 | 0.4287 |
| 0.1463 | 42.59 | 34500 | 3.5439 | 0.4270 |
| 0.139 | 44.44 | 36000 | 3.9381 | 0.4150 |
| 0.1352 | 46.3 | 37500 | 4.1744 | 0.4092 |
| 0.1369 | 48.15 | 39000 | 4.2279 | 0.4154 |
| 0.1273 | 50.0 | 40500 | 4.1691 | 0.4133 |
### Framework versions
- Transformers 4.16.0.dev0
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.11.0
|
ghadeermobasher/BC4CHEMD-Modified_BioM-ELECTRA-Base-Discriminator | 270cd4ef461fdedef10144e1e2ef10a4b6781939 | 2022-01-24T01:09:00.000Z | [
"pytorch",
"electra",
"token-classification",
"transformers",
"autotrain_compatible"
] | token-classification | false | ghadeermobasher | null | ghadeermobasher/BC4CHEMD-Modified_BioM-ELECTRA-Base-Discriminator | 1 | null | transformers | 29,004 | Entry not found |
ghadeermobasher/BC4CHEMD_ImbalancedBioM-ELECTRA-Base-Discriminator | 40302123f67c03baa52d789beb110e667eafd003 | 2022-01-23T09:09:55.000Z | [
"pytorch",
"electra",
"token-classification",
"transformers",
"autotrain_compatible"
] | token-classification | false | ghadeermobasher | null | ghadeermobasher/BC4CHEMD_ImbalancedBioM-ELECTRA-Base-Discriminator | 1 | null | transformers | 29,005 | Entry not found |
ghadeermobasher/BC5CDR-Disease_Modified_BioM-ELECTRA-Base-Discriminator | 6702c15b3c3d18356746d4f8cf08a4a2a7e44a2e | 2022-01-23T00:56:15.000Z | [
"pytorch",
"electra",
"token-classification",
"transformers",
"autotrain_compatible"
] | token-classification | false | ghadeermobasher | null | ghadeermobasher/BC5CDR-Disease_Modified_BioM-ELECTRA-Base-Discriminator | 1 | null | transformers | 29,006 | Entry not found |
ghadeermobasher/BioNLP13CG-Chem-Modified_BioM-ELECTRA-Base-Discriminator | ca855276d1afddfc64f70ec0ea0e99a0950d325f | 2022-01-22T23:23:20.000Z | [
"pytorch",
"electra",
"token-classification",
"transformers",
"autotrain_compatible"
] | token-classification | false | ghadeermobasher | null | ghadeermobasher/BioNLP13CG-Chem-Modified_BioM-ELECTRA-Base-Discriminator | 1 | null | transformers | 29,007 | Entry not found |
ghadeermobasher/CRAFT-Chem-Modified_BioM-ELECTRA-Base-Discriminator | eeaead148e7e04f8d4edb7922f676c4ea73307ff | 2022-01-23T02:10:50.000Z | [
"pytorch",
"electra",
"token-classification",
"transformers",
"autotrain_compatible"
] | token-classification | false | ghadeermobasher | null | ghadeermobasher/CRAFT-Chem-Modified_BioM-ELECTRA-Base-Discriminator | 1 | null | transformers | 29,008 | Entry not found |
ghhostboy/DialoGPT-medium-connorDBH3-1 | 7abfca2da8ed734e9cfbc6cc1489415e4f4975e4 | 2021-11-26T05:04:43.000Z | [
"pytorch",
"gpt2",
"text-generation",
"transformers",
"conversational"
] | conversational | false | ghhostboy | null | ghhostboy/DialoGPT-medium-connorDBH3-1 | 1 | null | transformers | 29,009 | ---
tags:
- conversational
---
# Connor |
ghofrani/common6 | eca7da668eb271998c14067f7e7c1928b21a8c99 | 2022-02-07T02:29:26.000Z | [
"pytorch",
"wav2vec2",
"automatic-speech-recognition",
"fa",
"dataset:common_voice",
"transformers",
"common_voice",
"generated_from_trainer",
"model-index"
] | automatic-speech-recognition | false | ghofrani | null | ghofrani/common6 | 1 | null | transformers | 29,010 | ---
language:
- fa
tags:
- automatic-speech-recognition
- common_voice
- generated_from_trainer
datasets:
- common_voice
model-index:
- name: common6
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. -->
# common6
This model is a fine-tuned version of [common6/checkpoint-3500](https://huggingface.co/common6/checkpoint-3500) on the COMMON_VOICE - FA dataset.
It achieves the following results on the evaluation set:
- Loss: 0.3706
- Wer: 0.3421
## 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: 6e-05
- train_batch_size: 32
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 256
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 100
- num_epochs: 200.0
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|:-------------:|:-----:|:-----:|:---------------:|:------:|
| 1.0344 | 10.0 | 500 | 0.4043 | 0.4511 |
| 0.9651 | 20.0 | 1000 | 0.3793 | 0.4159 |
| 0.9125 | 30.0 | 1500 | 0.3756 | 0.4046 |
| 0.8831 | 40.0 | 2000 | 0.3650 | 0.3876 |
| 0.8399 | 50.0 | 2500 | 0.3605 | 0.3772 |
| 0.819 | 60.0 | 3000 | 0.3622 | 0.3714 |
| 0.8029 | 70.0 | 3500 | 0.3561 | 0.3664 |
| 0.8104 | 80.0 | 4000 | 0.3595 | 0.3660 |
| 0.8118 | 90.0 | 4500 | 0.3460 | 0.3592 |
| 0.7831 | 100.0 | 5000 | 0.3566 | 0.3593 |
| 0.744 | 110.0 | 5500 | 0.3578 | 0.3535 |
| 0.7388 | 120.0 | 6000 | 0.3538 | 0.3520 |
| 0.714 | 130.0 | 6500 | 0.3682 | 0.3506 |
| 0.7291 | 140.0 | 7000 | 0.3625 | 0.3505 |
| 0.697 | 150.0 | 7500 | 0.3619 | 0.3479 |
| 0.6811 | 160.0 | 8000 | 0.3631 | 0.3440 |
| 0.6841 | 170.0 | 8500 | 0.3672 | 0.3460 |
| 0.6616 | 180.0 | 9000 | 0.3677 | 0.3410 |
| 0.6471 | 190.0 | 9500 | 0.3707 | 0.3420 |
| 0.6759 | 200.0 | 10000 | 0.3706 | 0.3421 |
### Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2
- Datasets 1.18.3.dev0
- Tokenizers 0.10.3
|
ghofrani/common7 | 9197cd4b03ac62816476789a661c7ccbafe25141 | 2022-02-04T01:32:24.000Z | [
"pytorch",
"wav2vec2",
"automatic-speech-recognition",
"fa",
"dataset:common_voice",
"transformers",
"mozilla-foundation/common_voice_7_0",
"generated_from_trainer",
"model-index"
] | automatic-speech-recognition | false | ghofrani | null | ghofrani/common7 | 1 | null | transformers | 29,011 | ---
language:
- fa
tags:
- automatic-speech-recognition
- mozilla-foundation/common_voice_7_0
- generated_from_trainer
datasets:
- common_voice
model-index:
- name: common7
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. -->
# common7
This model is a fine-tuned version of [common7/checkpoint-18500](https://huggingface.co/common7/checkpoint-18500) on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - FA dataset.
It achieves the following results on the evaluation set:
- Loss: 0.3448
- Wer: 0.3478
## 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: 6e-05
- train_batch_size: 32
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 100
- num_epochs: 150.0
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|:-------------:|:------:|:-----:|:---------------:|:------:|
| 2.957 | 3.29 | 500 | 2.9503 | 1.0 |
| 1.7225 | 6.58 | 1000 | 0.8860 | 0.7703 |
| 1.4907 | 9.86 | 1500 | 0.6555 | 0.6673 |
| 1.4177 | 13.16 | 2000 | 0.5784 | 0.6076 |
| 1.3425 | 16.45 | 2500 | 0.5379 | 0.5718 |
| 1.33 | 19.73 | 3000 | 0.4962 | 0.5245 |
| 1.4378 | 23.03 | 3500 | 0.4699 | 0.5098 |
| 1.1894 | 26.31 | 4000 | 0.4527 | 0.4848 |
| 1.1844 | 29.6 | 4500 | 0.4309 | 0.4651 |
| 1.1795 | 32.89 | 5000 | 0.4131 | 0.4524 |
| 1.1471 | 36.18 | 5500 | 0.4052 | 0.4435 |
| 1.1337 | 39.47 | 6000 | 0.3927 | 0.4363 |
| 1.1896 | 42.76 | 6500 | 0.3811 | 0.4254 |
| 1.1847 | 46.05 | 7000 | 0.3855 | 0.4129 |
| 0.9954 | 49.34 | 7500 | 0.3729 | 0.3981 |
| 1.0293 | 52.63 | 8000 | 0.3637 | 0.4014 |
| 1.0224 | 55.92 | 8500 | 0.3578 | 0.3885 |
| 1.012 | 59.21 | 9000 | 0.3629 | 0.3930 |
| 1.0772 | 62.5 | 9500 | 0.3635 | 0.3906 |
| 1.0344 | 65.79 | 10000 | 0.3469 | 0.3771 |
| 0.9457 | 69.08 | 10500 | 0.3435 | 0.3735 |
| 0.9307 | 72.37 | 11000 | 0.3519 | 0.3762 |
| 0.9523 | 75.65 | 11500 | 0.3443 | 0.3666 |
| 0.9523 | 78.94 | 12000 | 0.3502 | 0.3757 |
| 0.9475 | 82.24 | 12500 | 0.3509 | 0.3643 |
| 0.9971 | 85.52 | 13000 | 0.3502 | 0.3626 |
| 0.9058 | 88.81 | 13500 | 0.3472 | 0.3605 |
| 0.8922 | 92.1 | 14000 | 0.3530 | 0.3618 |
| 0.9 | 95.39 | 14500 | 0.3500 | 0.3574 |
| 0.9051 | 98.68 | 15000 | 0.3456 | 0.3535 |
| 0.9304 | 101.97 | 15500 | 0.3438 | 0.3578 |
| 0.9433 | 105.26 | 16000 | 0.3396 | 0.3530 |
| 0.8988 | 108.55 | 16500 | 0.3436 | 0.3539 |
| 0.8789 | 111.84 | 17000 | 0.3426 | 0.3516 |
| 0.8667 | 115.13 | 17500 | 0.3438 | 0.3506 |
| 0.8895 | 118.42 | 18000 | 0.3434 | 0.3503 |
| 0.8888 | 121.71 | 18500 | 0.3425 | 0.3494 |
| 0.9453 | 125.0 | 19000 | 0.3415 | 0.3480 |
| 0.9267 | 128.29 | 19500 | 0.3477 | 0.3503 |
| 0.8315 | 131.58 | 20000 | 0.3476 | 0.3505 |
| 0.8542 | 134.86 | 20500 | 0.3475 | 0.3506 |
| 0.8478 | 138.16 | 21000 | 0.3430 | 0.3481 |
| 0.8643 | 141.45 | 21500 | 0.3451 | 0.3485 |
| 0.8705 | 144.73 | 22000 | 0.3444 | 0.3474 |
| 0.9869 | 148.03 | 22500 | 0.3441 | 0.3493 |
### Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2
- Datasets 1.18.3.dev0
- Tokenizers 0.10.3
|
ghofrani/common8 | c932f110d73bc6a59184a42b0772b9c4f9ad93d0 | 2022-02-08T23:51:46.000Z | [
"pytorch",
"wav2vec2",
"automatic-speech-recognition",
"fa",
"dataset:common_voice",
"transformers",
"mozilla-foundation/common_voice_8_0",
"generated_from_trainer",
"model-index"
] | automatic-speech-recognition | false | ghofrani | null | ghofrani/common8 | 1 | null | transformers | 29,012 | ---
language:
- fa
tags:
- automatic-speech-recognition
- mozilla-foundation/common_voice_8_0
- generated_from_trainer
datasets:
- common_voice
model-index:
- name: common8
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. -->
# common8
This model is a fine-tuned version of [wghts/checkpoint-20000](https://huggingface.co/wghts/checkpoint-20000) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - FA dataset.
It achieves the following results on the evaluation set:
- Loss: 0.3174
- Wer: 0.3022
## 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: 1e-06
- train_batch_size: 32
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 6
- total_train_batch_size: 192
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 100
- num_epochs: 250.0
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|:-------------:|:------:|:-----:|:---------------:|:------:|
| 3.5847 | 1.93 | 500 | 3.5104 | 1.0 |
| 2.7858 | 3.86 | 1000 | 2.9601 | 1.0001 |
| 1.6827 | 5.79 | 1500 | 0.7853 | 0.7030 |
| 1.4656 | 7.72 | 2000 | 0.6076 | 0.6014 |
| 1.3693 | 9.65 | 2500 | 0.5114 | 0.5307 |
| 1.379 | 11.58 | 3000 | 0.4666 | 0.4940 |
| 1.2832 | 13.51 | 3500 | 0.4257 | 0.4593 |
| 1.1931 | 15.44 | 4000 | 0.4039 | 0.4427 |
| 1.2911 | 17.37 | 4500 | 0.3956 | 0.4295 |
| 1.1577 | 19.3 | 5000 | 0.3705 | 0.4114 |
| 1.1135 | 21.24 | 5500 | 0.3740 | 0.4010 |
| 1.19 | 23.17 | 6000 | 0.3611 | 0.3935 |
| 1.1008 | 25.1 | 6500 | 0.3503 | 0.3880 |
| 1.0805 | 27.03 | 7000 | 0.3427 | 0.3781 |
| 1.1556 | 28.96 | 7500 | 0.3442 | 0.3727 |
| 1.0596 | 30.89 | 8000 | 0.3398 | 0.3646 |
| 1.0219 | 32.82 | 8500 | 0.3312 | 0.3660 |
| 1.1042 | 34.75 | 9000 | 0.3287 | 0.3612 |
| 1.0273 | 36.68 | 9500 | 0.3236 | 0.3556 |
| 1.0383 | 38.61 | 10000 | 0.3217 | 0.3558 |
| 1.0498 | 40.54 | 10500 | 0.3205 | 0.3520 |
| 0.9969 | 42.47 | 11000 | 0.3125 | 0.3504 |
| 1.0658 | 44.4 | 11500 | 0.3120 | 0.3493 |
| 0.992 | 46.33 | 12000 | 0.3137 | 0.3476 |
| 0.9737 | 48.26 | 12500 | 0.3085 | 0.3413 |
| 1.0817 | 50.19 | 13000 | 0.3091 | 0.3418 |
| 0.9414 | 52.12 | 13500 | 0.3072 | 0.3344 |
| 0.9295 | 54.05 | 14000 | 0.3039 | 0.3322 |
| 1.0248 | 55.98 | 14500 | 0.2991 | 0.3325 |
| 0.9474 | 57.91 | 15000 | 0.3032 | 0.3348 |
| 0.928 | 59.85 | 15500 | 0.2999 | 0.3285 |
| 1.0321 | 61.78 | 16000 | 0.2982 | 0.3253 |
| 0.9255 | 63.71 | 16500 | 0.2970 | 0.3231 |
| 0.8928 | 65.64 | 17000 | 0.2993 | 0.3250 |
| 1.008 | 67.57 | 17500 | 0.2985 | 0.3222 |
| 0.9371 | 69.5 | 18000 | 0.2968 | 0.3216 |
| 0.9077 | 71.43 | 18500 | 0.3011 | 0.3299 |
| 1.0044 | 73.36 | 19000 | 0.3053 | 0.3306 |
| 0.9625 | 75.29 | 19500 | 0.3159 | 0.3295 |
| 0.9816 | 77.22 | 20000 | 0.3080 | 0.3304 |
| 0.9587 | 119.19 | 20500 | 0.3088 | 0.3284 |
| 0.9178 | 122.09 | 21000 | 0.3132 | 0.3320 |
| 1.0282 | 125.0 | 21500 | 0.3099 | 0.3266 |
| 0.9337 | 127.9 | 22000 | 0.3110 | 0.3317 |
| 0.8822 | 130.81 | 22500 | 0.3037 | 0.3247 |
| 0.9644 | 133.72 | 23000 | 0.3037 | 0.3238 |
| 0.9214 | 136.62 | 23500 | 0.3040 | 0.3234 |
| 0.9167 | 139.53 | 24000 | 0.3079 | 0.3203 |
| 0.9047 | 142.44 | 24500 | 0.3018 | 0.3177 |
| 0.8909 | 145.35 | 25000 | 0.3053 | 0.3181 |
| 0.9646 | 148.25 | 25500 | 0.3095 | 0.3229 |
| 0.8802 | 151.16 | 26000 | 0.3111 | 0.3192 |
| 0.8411 | 154.07 | 26500 | 0.3068 | 0.3123 |
| 0.9235 | 156.97 | 27000 | 0.3090 | 0.3177 |
| 0.8943 | 159.88 | 27500 | 0.3115 | 0.3179 |
| 0.8854 | 162.79 | 28000 | 0.3052 | 0.3157 |
| 0.8734 | 165.69 | 28500 | 0.3077 | 0.3124 |
| 0.8515 | 168.6 | 29000 | 0.3117 | 0.3128 |
| 0.912 | 171.51 | 29500 | 0.3039 | 0.3121 |
| 0.8669 | 174.42 | 30000 | 0.3120 | 0.3123 |
| 0.823 | 177.32 | 30500 | 0.3148 | 0.3118 |
| 0.9129 | 180.23 | 31000 | 0.3179 | 0.3101 |
| 0.8255 | 183.14 | 31500 | 0.3164 | 0.3114 |
| 0.8948 | 186.05 | 32000 | 0.3128 | 0.3101 |
| 0.8397 | 188.95 | 32500 | 0.3143 | 0.3068 |
| 0.8341 | 191.86 | 33000 | 0.3127 | 0.3136 |
| 0.873 | 194.76 | 33500 | 0.3149 | 0.3124 |
| 0.8232 | 197.67 | 34000 | 0.3166 | 0.3086 |
| 0.8002 | 200.58 | 34500 | 0.3149 | 0.3061 |
| 0.8621 | 203.49 | 35000 | 0.3160 | 0.3093 |
| 0.8123 | 206.39 | 35500 | 0.3141 | 0.3063 |
| 0.7995 | 209.3 | 36000 | 0.3174 | 0.3075 |
| 0.8271 | 212.21 | 36500 | 0.3173 | 0.3043 |
| 0.8059 | 215.12 | 37000 | 0.3176 | 0.3079 |
| 0.8835 | 218.02 | 37500 | 0.3169 | 0.3062 |
| 0.8027 | 220.93 | 38000 | 0.3203 | 0.3098 |
| 0.775 | 223.83 | 38500 | 0.3159 | 0.3068 |
| 0.8487 | 226.74 | 39000 | 0.3161 | 0.3072 |
| 0.7929 | 229.65 | 39500 | 0.3143 | 0.3037 |
| 0.7653 | 232.56 | 40000 | 0.3160 | 0.3048 |
| 0.8211 | 235.46 | 40500 | 0.3173 | 0.3031 |
| 0.7761 | 238.37 | 41000 | 0.3176 | 0.3025 |
| 0.7761 | 241.28 | 41500 | 0.3179 | 0.3027 |
| 0.7903 | 244.19 | 42000 | 0.3181 | 0.3016 |
| 0.7807 | 247.09 | 42500 | 0.3170 | 0.3027 |
| 0.8406 | 250.0 | 43000 | 0.3174 | 0.3022 |
### Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2
- Datasets 1.18.3.dev0
- Tokenizers 0.10.3
|
ghofrani/wghts | fbee6a05cbce5b6e140cd0d2228ddb44c98a919b | 2022-02-01T14:40:53.000Z | [
"pytorch",
"wav2vec2",
"automatic-speech-recognition",
"transformers"
] | automatic-speech-recognition | false | ghofrani | null | ghofrani/wghts | 1 | null | transformers | 29,013 | Entry not found |
ghofrani/xls-r-300m-fa | cd91fbf24b356c04eb76bbf2902371be0332ff4a | 2022-01-30T17:57:19.000Z | [
"pytorch",
"wav2vec2",
"automatic-speech-recognition",
"transformers"
] | automatic-speech-recognition | false | ghofrani | null | ghofrani/xls-r-300m-fa | 1 | null | transformers | 29,014 | Entry not found |
giganticode/bert-large-StackOverflow-comments_1M | 3d001537561adbeaf5eceea41652045e2e05bdad | 2021-10-25T13:05:34.000Z | [
"pytorch",
"bert",
"fill-mask",
"transformers",
"autotrain_compatible"
] | fill-mask | false | giganticode | null | giganticode/bert-large-StackOverflow-comments_1M | 1 | null | transformers | 29,015 | Entry not found |
glasses/deit_tiny_patch16_224 | 644ed2bb334f3134610ed3e322f5f851d0151aff | 2021-04-22T18:44:18.000Z | [
"pytorch",
"arxiv:2010.11929",
"transformers"
] | null | false | glasses | null | glasses/deit_tiny_patch16_224 | 1 | null | transformers | 29,016 | # deit_tiny_patch16_224
Implementation of DeiT proposed in [Training data-efficient image
transformers & distillation through
attention](https://arxiv.org/pdf/2010.11929.pdf)
An attention based distillation is proposed where a new token is added
to the model, the [dist]{.title-ref} token.
``` {.sourceCode .}
DeiT.deit_tiny_patch16_224()
DeiT.deit_small_patch16_224()
DeiT.deit_base_patch16_224()
DeiT.deit_base_patch16_384()
```
|
glasses/densenet201 | 06d094765920f89ade917d7543408af6aeec1400 | 2021-12-01T07:49:34.000Z | [
"pytorch",
"arxiv:1608.06993",
"transformers"
] | null | false | glasses | null | glasses/densenet201 | 1 | null | transformers | 29,017 | # densenet201
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/efficientnet_b2 | a7e47f80d02a476ad3c1bb93a6dbf0f9ed3b0e87 | 2021-12-01T08:08:06.000Z | [
"pytorch",
"arxiv:1905.11946",
"transformers"
] | null | false | glasses | null | glasses/efficientnet_b2 | 1 | null | transformers | 29,018 | # efficientnet_b2
Implementation of EfficientNet proposed in [EfficientNet: Rethinking
Model Scaling for Convolutional Neural
Networks](https://arxiv.org/abs/1905.11946)
The basic architecture is similar to MobileNetV2 as was computed by
using [Progressive Neural Architecture
Search](https://arxiv.org/abs/1905.11946) .
The following table shows the basic architecture
(EfficientNet-efficientnet\_b0):
Then, the architecture is scaled up from
[-efficientnet\_b0]{.title-ref} to [-efficientnet\_b7]{.title-ref}
using compound scaling.
``` python
EfficientNet.efficientnet_b0()
EfficientNet.efficientnet_b1()
EfficientNet.efficientnet_b2()
EfficientNet.efficientnet_b3()
EfficientNet.efficientnet_b4()
EfficientNet.efficientnet_b5()
EfficientNet.efficientnet_b6()
EfficientNet.efficientnet_b7()
EfficientNet.efficientnet_b8()
EfficientNet.efficientnet_l2()
```
Examples:
``` python
EfficientNet.efficientnet_b0(activation = nn.SELU)
# change number of classes (default is 1000 )
EfficientNet.efficientnet_b0(n_classes=100)
# pass a different block
EfficientNet.efficientnet_b0(block=...)
# store each feature
x = torch.rand((1, 3, 224, 224))
model = EfficientNet.efficientnet_b0()
# 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, 32, 112, 112]), torch.Size([1, 24, 56, 56]), torch.Size([1, 40, 28, 28]), torch.Size([1, 80, 14, 14])]
```
|
glasses/efficientnet_lite0 | 053175a3c8d6b541c9c74a1824e204022c939267 | 2021-12-01T08:09:01.000Z | [
"pytorch",
"transformers"
] | null | false | glasses | null | glasses/efficientnet_lite0 | 1 | null | transformers | 29,019 | Entry not found |
glasses/regnetx_002 | 95112ce3a26a6732f545f89845acfffd10b40dc4 | 2021-11-30T20:25:54.000Z | [
"pytorch",
"arxiv:2003.13678",
"transformers"
] | null | false | glasses | null | glasses/regnetx_002 | 1 | null | transformers | 29,020 | # regnetx_002
Implementation of RegNet proposed in [Designing Network Design
Spaces](https://arxiv.org/abs/2003.13678)
The main idea is to start with a high dimensional search space and
iteratively reduce the search space by empirically apply constrains
based on the best performing models sampled by the current search
space.
The resulting models are light, accurate, and faster than
EfficientNets (up to 5x times!)
For example, to go from $AnyNet_A$ to $AnyNet_B$ they fixed the
bottleneck ratio $b_i$ for all stage $i$. The following table shows
all the restrictions applied from one search space to the next one.
The paper is really well written and very interesting, I highly
recommended read it.
``` python
ResNet.regnetx_002()
ResNet.regnetx_004()
ResNet.regnetx_006()
ResNet.regnetx_008()
ResNet.regnetx_016()
ResNet.regnetx_040()
ResNet.regnetx_064()
ResNet.regnetx_080()
ResNet.regnetx_120()
ResNet.regnetx_160()
ResNet.regnetx_320()
# Y variants (with SE)
ResNet.regnety_002()
# ...
ResNet.regnetx_320()
You can easily customize your model
```
Examples:
``` python
# change activation
RegNet.regnetx_004(activation = nn.SELU)
# change number of classes (default is 1000 )
RegNet.regnetx_004(n_classes=100)
# pass a different block
RegNet.regnetx_004(block=RegNetYBotteneckBlock)
# change the steam
model = RegNet.regnetx_004(stem=ResNetStemC)
change shortcut
model = RegNet.regnetx_004(block=partial(RegNetYBotteneckBlock, shortcut=ResNetShorcutD))
# store each feature
x = torch.rand((1, 3, 224, 224))
# get features
model = RegNet.regnetx_004()
# 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, 32, 112, 112]), torch.Size([1, 32, 56, 56]), torch.Size([1, 64, 28, 28]), torch.Size([1, 160, 14, 14])]
```
|
glasses/regnetx_040 | ae9ad4f198ee581a5fabb74e1c12ab76aa4be8f3 | 2021-11-30T20:27:44.000Z | [
"pytorch",
"transformers"
] | null | false | glasses | null | glasses/regnetx_040 | 1 | null | transformers | 29,021 | Entry not found |
glasses/regnety_002 | e329d1379bc33fb17896af32a7b58cb8aaf4857c | 2021-12-01T07:45:22.000Z | [
"pytorch",
"arxiv:2003.13678",
"transformers"
] | null | false | glasses | null | glasses/regnety_002 | 1 | null | transformers | 29,022 | # regnety_002
Implementation of RegNet proposed in [Designing Network Design
Spaces](https://arxiv.org/abs/2003.13678)
The main idea is to start with a high dimensional search space and
iteratively reduce the search space by empirically apply constrains
based on the best performing models sampled by the current search
space.
The resulting models are light, accurate, and faster than
EfficientNets (up to 5x times!)
For example, to go from $AnyNet_A$ to $AnyNet_B$ they fixed the
bottleneck ratio $b_i$ for all stage $i$. The following table shows
all the restrictions applied from one search space to the next one.
The paper is really well written and very interesting, I highly
recommended read it.
``` python
ResNet.regnetx_002()
ResNet.regnetx_004()
ResNet.regnetx_006()
ResNet.regnetx_008()
ResNet.regnetx_016()
ResNet.regnetx_040()
ResNet.regnetx_064()
ResNet.regnetx_080()
ResNet.regnetx_120()
ResNet.regnetx_160()
ResNet.regnetx_320()
# Y variants (with SE)
ResNet.regnety_002()
# ...
ResNet.regnetx_320()
You can easily customize your model
```
Examples:
``` python
# change activation
RegNet.regnetx_004(activation = nn.SELU)
# change number of classes (default is 1000 )
RegNet.regnetx_004(n_classes=100)
# pass a different block
RegNet.regnetx_004(block=RegNetYBotteneckBlock)
# change the steam
model = RegNet.regnetx_004(stem=ResNetStemC)
change shortcut
model = RegNet.regnetx_004(block=partial(RegNetYBotteneckBlock, shortcut=ResNetShorcutD))
# store each feature
x = torch.rand((1, 3, 224, 224))
# get features
model = RegNet.regnetx_004()
# 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, 32, 112, 112]), torch.Size([1, 32, 56, 56]), torch.Size([1, 64, 28, 28]), torch.Size([1, 160, 14, 14])]
```
|
glasses/regnety_008 | c11a1def96cd4792b8aac07fc5a56bb348a2f8e1 | 2021-12-01T07:46:29.000Z | [
"pytorch",
"arxiv:2003.13678",
"transformers"
] | null | false | glasses | null | glasses/regnety_008 | 1 | null | transformers | 29,023 | # regnety_008
Implementation of RegNet proposed in [Designing Network Design
Spaces](https://arxiv.org/abs/2003.13678)
The main idea is to start with a high dimensional search space and
iteratively reduce the search space by empirically apply constrains
based on the best performing models sampled by the current search
space.
The resulting models are light, accurate, and faster than
EfficientNets (up to 5x times!)
For example, to go from $AnyNet_A$ to $AnyNet_B$ they fixed the
bottleneck ratio $b_i$ for all stage $i$. The following table shows
all the restrictions applied from one search space to the next one.
The paper is really well written and very interesting, I highly
recommended read it.
``` python
ResNet.regnetx_002()
ResNet.regnetx_004()
ResNet.regnetx_006()
ResNet.regnetx_008()
ResNet.regnetx_016()
ResNet.regnetx_040()
ResNet.regnetx_064()
ResNet.regnetx_080()
ResNet.regnetx_120()
ResNet.regnetx_160()
ResNet.regnetx_320()
# Y variants (with SE)
ResNet.regnety_002()
# ...
ResNet.regnetx_320()
You can easily customize your model
```
Examples:
``` python
# change activation
RegNet.regnetx_004(activation = nn.SELU)
# change number of classes (default is 1000 )
RegNet.regnetx_004(n_classes=100)
# pass a different block
RegNet.regnetx_004(block=RegNetYBotteneckBlock)
# change the steam
model = RegNet.regnetx_004(stem=ResNetStemC)
change shortcut
model = RegNet.regnetx_004(block=partial(RegNetYBotteneckBlock, shortcut=ResNetShorcutD))
# store each feature
x = torch.rand((1, 3, 224, 224))
# get features
model = RegNet.regnetx_004()
# 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, 32, 112, 112]), torch.Size([1, 32, 56, 56]), torch.Size([1, 64, 28, 28]), torch.Size([1, 160, 14, 14])]
```
|
glasses/regnety_032 | f81cec2a53473c19368a04807787b598b6761e5b | 2021-12-01T08:31:57.000Z | [
"pytorch",
"transformers"
] | null | false | glasses | null | glasses/regnety_032 | 1 | null | transformers | 29,024 | Entry not found |
glasses/regnety_064 | 9264ebae2f90665208a23f49cf3810bd1e56c504 | 2021-12-01T07:48:05.000Z | [
"pytorch",
"transformers"
] | null | false | glasses | null | glasses/regnety_064 | 1 | null | transformers | 29,025 | Entry not found |
glasses/resnet101 | 34e388c831995fd4fdde057d194a0cc4e8dece76 | 2021-11-30T20:10:49.000Z | [
"pytorch",
"transformers"
] | null | false | glasses | null | glasses/resnet101 | 1 | null | transformers | 29,026 | Entry not found |
glasses/resnet26d | 4256019677b10c45bdb178090012fe7a4287042a | 2021-11-30T20:07:33.000Z | [
"pytorch",
"dataset:imagenet",
"arxiv:1512.03385",
"arxiv:1812.01187",
"transformers",
"image-classification",
"license:apache-2.0"
] | image-classification | false | glasses | null | glasses/resnet26d | 1 | null | transformers | 29,027 | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# resnet26d
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/vgg11 | 705c39d6a90328422a3ffac6e0399f07ba46ca29 | 2021-12-01T07:53:25.000Z | [
"pytorch",
"transformers"
] | null | false | glasses | null | glasses/vgg11 | 1 | null | transformers | 29,028 | # vgg11
Implementation of VGG proposed in [Very Deep Convolutional Networks For
Large-Scale Image Recognition](https://arxiv.org/pdf/1409.1556.pdf)
``` python
VGG.vgg11()
VGG.vgg13()
VGG.vgg16()
VGG.vgg19()
VGG.vgg11_bn()
VGG.vgg13_bn()
VGG.vgg16_bn()
VGG.vgg19_bn()
```
Please be aware that the [bn]{.title-ref} models uses BatchNorm but
they are very old and people back then don\'t know the bias is
superfluous in a conv followed by a batchnorm.
Examples:
``` python
# change activation
VGG.vgg11(activation = nn.SELU)
# change number of classes (default is 1000 )
VGG.vgg11(n_classes=100)
# pass a different block
from nn.models.classification.senet import SENetBasicBlock
VGG.vgg11(block=SENetBasicBlock)
# store the features tensor after every block
```
|
glasses/vit_base_patch16_224 | b40b132218b4bc9b8ca2e440a6dec7c9f03b3ff1 | 2021-12-01T08:23:58.000Z | [
"pytorch",
"arxiv:2010.11929",
"transformers"
] | null | false | glasses | null | glasses/vit_base_patch16_224 | 1 | null | transformers | 29,029 | # vit_base_patch16_224
Implementation of Vision Transformer (ViT) proposed in [An Image Is
Worth 16x16 Words: Transformers For Image Recognition At
Scale](https://arxiv.org/pdf/2010.11929.pdf)
The following image from the authors shows the architecture.
``` python
ViT.vit_small_patch16_224()
ViT.vit_base_patch16_224()
ViT.vit_base_patch16_384()
ViT.vit_base_patch32_384()
ViT.vit_huge_patch16_224()
ViT.vit_huge_patch32_384()
ViT.vit_large_patch16_224()
ViT.vit_large_patch16_384()
ViT.vit_large_patch32_384()
```
Examples:
``` python
# change activation
ViT.vit_base_patch16_224(activation = nn.SELU)
# change number of classes (default is 1000 )
ViT.vit_base_patch16_224(n_classes=100)
# pass a different block, default is TransformerEncoderBlock
ViT.vit_base_patch16_224(block=MyCoolTransformerBlock)
# get features
model = ViT.vit_base_patch16_224
# 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, 197, 768]), torch.Size([1, 197, 768]), ...]
# change the tokens, you have to subclass ViTTokens
class MyTokens(ViTTokens):
def __init__(self, emb_size: int):
super().__init__(emb_size)
self.my_new_token = nn.Parameter(torch.randn(1, 1, emb_size))
ViT(tokens=MyTokens)
```
|
glasses/vit_huge_patch16_224 | 3fe73ba563e88fb761eb9602963b7ac89a09e430 | 2021-04-22T18:39:36.000Z | [
"pytorch",
"arxiv:2010.11929",
"transformers"
] | null | false | glasses | null | glasses/vit_huge_patch16_224 | 1 | null | transformers | 29,030 | # vit_huge_patch16_224
Implementation of Vision Transformer (ViT) proposed in [An Image Is
Worth 16x16 Words: Transformers For Image Recognition At
Scale](https://arxiv.org/pdf/2010.11929.pdf)
The following image from the authors shows the architecture.
``` python
ViT.vit_small_patch16_224()
ViT.vit_base_patch16_224()
ViT.vit_base_patch16_384()
ViT.vit_base_patch32_384()
ViT.vit_huge_patch16_224()
ViT.vit_huge_patch32_384()
ViT.vit_large_patch16_224()
ViT.vit_large_patch16_384()
ViT.vit_large_patch32_384()
```
Examples:
``` python
# change activation
ViT.vit_base_patch16_224(activation = nn.SELU)
# change number of classes (default is 1000 )
ViT.vit_base_patch16_224(n_classes=100)
# pass a different block, default is TransformerEncoderBlock
ViT.vit_base_patch16_224(block=MyCoolTransformerBlock)
# get features
model = ViT.vit_base_patch16_224
# 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, 197, 768]), torch.Size([1, 197, 768]), ...]
# change the tokens, you have to subclass ViTTokens
class MyTokens(ViTTokens):
def __init__(self, emb_size: int):
super().__init__(emb_size)
self.my_new_token = nn.Parameter(torch.randn(1, 1, emb_size))
ViT(tokens=MyTokens)
```
|
glasses/vit_huge_patch32_384 | 5fe4bb24cddf4b5bdb69ed43e6e79b0cc584d65e | 2021-04-22T18:41:37.000Z | [
"pytorch",
"arxiv:2010.11929",
"transformers"
] | null | false | glasses | null | glasses/vit_huge_patch32_384 | 1 | null | transformers | 29,031 | # vit_huge_patch32_384
Implementation of Vision Transformer (ViT) proposed in [An Image Is
Worth 16x16 Words: Transformers For Image Recognition At
Scale](https://arxiv.org/pdf/2010.11929.pdf)
The following image from the authors shows the architecture.
``` python
ViT.vit_small_patch16_224()
ViT.vit_base_patch16_224()
ViT.vit_base_patch16_384()
ViT.vit_base_patch32_384()
ViT.vit_huge_patch16_224()
ViT.vit_huge_patch32_384()
ViT.vit_large_patch16_224()
ViT.vit_large_patch16_384()
ViT.vit_large_patch32_384()
```
Examples:
``` python
# change activation
ViT.vit_base_patch16_224(activation = nn.SELU)
# change number of classes (default is 1000 )
ViT.vit_base_patch16_224(n_classes=100)
# pass a different block, default is TransformerEncoderBlock
ViT.vit_base_patch16_224(block=MyCoolTransformerBlock)
# get features
model = ViT.vit_base_patch16_224
# 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, 197, 768]), torch.Size([1, 197, 768]), ...]
# change the tokens, you have to subclass ViTTokens
class MyTokens(ViTTokens):
def __init__(self, emb_size: int):
super().__init__(emb_size)
self.my_new_token = nn.Parameter(torch.randn(1, 1, emb_size))
ViT(tokens=MyTokens)
```
|
gniemiec/mt5-small-finetuned-xsum | 1b7410ef252bdd00a331c8e7bd8dae2b1a0e0d19 | 2021-09-21T13:22:57.000Z | [
"pytorch",
"tensorboard",
"mt5",
"text2text-generation",
"dataset:xsum",
"transformers",
"generated_from_trainer",
"license:apache-2.0",
"model-index",
"autotrain_compatible"
] | text2text-generation | false | gniemiec | null | gniemiec/mt5-small-finetuned-xsum | 1 | null | transformers | 29,032 | ---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- xsum
metrics:
- rouge
model-index:
- name: mt5-small-finetuned-xsum
results:
- task:
name: Sequence-to-sequence Language Modeling
type: text2text-generation
dataset:
name: xsum
type: xsum
args: default
metrics:
- name: Rouge1
type: rouge
value: 2.8351
---
<!-- 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-small-finetuned-xsum
This model is a fine-tuned version of [google/mt5-small](https://huggingface.co/google/mt5-small) on the xsum dataset.
It achieves the following results on the evaluation set:
- Loss: nan
- Rouge1: 2.8351
- Rouge2: 0.3143
- Rougel: 2.6488
- Rougelsum: 2.6463
- Gen Len: 4.9416
## 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
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Validation Loss | Rouge1 | Rouge2 | Rougel | Rougelsum | Gen Len |
|:-------------:|:-----:|:-----:|:---------------:|:------:|:------:|:------:|:---------:|:-------:|
| nan | 1.0 | 12753 | nan | 2.8351 | 0.3143 | 2.6488 | 2.6463 | 4.9416 |
### Framework versions
- Transformers 4.10.2
- Pytorch 1.9.0+cu102
- Datasets 1.12.1
- Tokenizers 0.10.3
|
gonced8/pegasus-qa | ea2408de301ae6f79c879572bf9569f57c24551b | 2021-05-17T15:46:42.000Z | [
"pytorch",
"pegasus",
"text2text-generation",
"transformers",
"autotrain_compatible"
] | text2text-generation | false | gonced8 | null | gonced8/pegasus-qa | 1 | null | transformers | 29,033 | Entry not found |
goodjw/gpt-trinity-poem | 6aeec8ffb04bde11334d1c61be1061f685ee7d4f | 2021-12-07T15:38:01.000Z | [
"pytorch",
"gpt2",
"text-generation",
"transformers"
] | text-generation | false | goodjw | null | goodjw/gpt-trinity-poem | 1 | null | transformers | 29,034 | Entry not found |
google/multiberts-seed_0-step_0k | 93e2ee33230f548b15bebc8c936f0cb5ef17b37a | 2021-11-05T23:37:18.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_0k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_0k | 1 | null | transformers | 29,035 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_0k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, 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 #0, 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_0-step_0k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-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_0-step_0k')
model = BertModel.from_pretrained("google/multiberts-seed_0-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_0-step_1000k | 6d16dba114c3c91aecee3526ffbaac84f7f5e90f | 2021-11-06T00:08:04.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_1000k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_1000k | 1 | null | transformers | 29,036 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_1000k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 1000k
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 #0, captured at step 1000k (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_0-step_1000k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_1000k")
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_0-step_1000k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_1000k")
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_0-step_100k | 144b03be7bcb188b4c32b1b1d1b4588e3492de07 | 2021-11-05T23:45:48.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_100k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_100k | 1 | null | transformers | 29,037 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_100k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 100k
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 #0, captured at step 100k (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_0-step_100k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_100k")
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_0-step_100k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_100k")
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_0-step_1100k | f33754924915b9b4adabb88c9a0258fddd4cbe48 | 2021-11-06T00:09:45.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_1100k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_1100k | 1 | null | transformers | 29,038 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_1100k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, 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 #0, 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_0-step_1100k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-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_0-step_1100k')
model = BertModel.from_pretrained("google/multiberts-seed_0-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_0-step_1300k | f084998a12cb66a1791644c2cf77695daba9bd1f | 2021-11-06T00:13:16.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_1300k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_1300k | 1 | null | transformers | 29,039 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_1300k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 1300k
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 #0, captured at step 1300k (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_0-step_1300k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_1300k")
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_0-step_1300k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_1300k")
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_0-step_1600k | 00d5c4d9a1a5711b6a508db1110c666cd6596e1d | 2021-11-06T00:18:39.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_1600k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_1600k | 1 | null | transformers | 29,040 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_1600k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 1600k
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 #0, captured at step 1600k (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_0-step_1600k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_1600k")
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_0-step_1600k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_1600k")
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_0-step_1800k | f813aa4d9f8d566efaf4e400383ed20684d4bfe5 | 2021-11-06T00:22:02.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_1800k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_1800k | 1 | null | transformers | 29,041 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_1800k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, 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 #0, 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_0-step_1800k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-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_0-step_1800k')
model = BertModel.from_pretrained("google/multiberts-seed_0-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_0-step_180k | 0419b80e6129e38ecf0bf4380daa01de5055dbf3 | 2021-11-05T23:52:40.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_180k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_180k | 1 | null | transformers | 29,042 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_180k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 180k
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 #0, captured at step 180k (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_0-step_180k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_180k")
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_0-step_180k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_180k")
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_0-step_1900k | c6e05e75cb669fe8874deec83d3145a5203dd18c | 2021-11-06T00:23:42.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_1900k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_1900k | 1 | null | transformers | 29,043 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_1900k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 1900k
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 #0, captured at step 1900k (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_0-step_1900k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_1900k")
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_0-step_1900k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_1900k")
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_0-step_20k | 8c40009a439f3e7f9c5844f1707a4a1c46197b99 | 2021-11-05T23:38:42.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_20k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_20k | 1 | null | transformers | 29,044 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_20k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 20k
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 #0, captured at step 20k (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_0-step_20k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_20k")
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_0-step_20k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_20k")
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_0-step_40k | 49397d8a5abe7a220e7adc5d9e10ded4c44f9c9b | 2021-11-05T23:40:21.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_40k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_40k | 1 | null | transformers | 29,045 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_40k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, 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 #0, 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_0-step_40k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-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_0-step_40k')
model = BertModel.from_pretrained("google/multiberts-seed_0-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_0-step_500k | 7a3142802965daf1d830a8d7ef20fc07bb52a388 | 2021-11-05T23:59:35.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_500k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_500k | 1 | null | transformers | 29,046 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_500k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, 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 #0, 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_0-step_500k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-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_0-step_500k')
model = BertModel.from_pretrained("google/multiberts-seed_0-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_0-step_600k | 49bd26b107a6edfa4ce67849cc7ec0d7316cd586 | 2021-11-06T00:01:17.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_600k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_600k | 1 | null | transformers | 29,047 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_600k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 600k
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 #0, captured at step 600k (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_0-step_600k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_600k")
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_0-step_600k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_600k")
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_0-step_60k | 9ab74d84068694cf164abb169c44cb647421caf4 | 2021-11-05T23:42:02.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_60k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_60k | 1 | null | transformers | 29,048 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_60k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, Step 60k
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 #0, captured at step 60k (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_0-step_60k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-step_60k")
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_0-step_60k')
model = BertModel.from_pretrained("google/multiberts-seed_0-step_60k")
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_0-step_80k | f9acf7e8fa6c803090b5cea8607ba4070bdc5d90 | 2021-11-05T23:44:05.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_80k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_80k | 1 | null | transformers | 29,049 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_80k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, 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 #0, 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_0-step_80k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-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_0-step_80k')
model = BertModel.from_pretrained("google/multiberts-seed_0-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_0-step_900k | 08a690ae07a6f913cc6cb17f29868c9c4dd166ee | 2021-11-06T00:06:27.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_0",
"multiberts-seed_0-step_900k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_0-step_900k | 1 | null | transformers | 29,050 | ---
language: en
tags:
- multiberts
- multiberts-seed_0
- multiberts-seed_0-step_900k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 0, 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 #0, 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_0-step_900k')
model = TFBertModel.from_pretrained("google/multiberts-seed_0-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_0-step_900k')
model = BertModel.from_pretrained("google/multiberts-seed_0-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/multiberts-seed_1-step_0k | 6461363270fb78b7322d003717392f2d37cee00f | 2021-11-06T00:33:14.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_0k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_0k | 1 | null | transformers | 29,051 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_0k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, 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 #1, 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_1-step_0k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-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_1-step_0k')
model = BertModel.from_pretrained("google/multiberts-seed_1-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_1-step_1200k | 3a77ec97c76fc286f29ade9f215502f02650fcf5 | 2021-11-06T01:06:43.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_1200k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_1200k | 1 | null | transformers | 29,052 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_1200k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, Step 1200k
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 #1, captured at step 1200k (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_1-step_1200k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-step_1200k")
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_1-step_1200k')
model = BertModel.from_pretrained("google/multiberts-seed_1-step_1200k")
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_1-step_140k | 499f45fb52ea6add6e78d94fff7e266badfaed97 | 2021-11-06T00:45:09.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_140k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_140k | 1 | null | transformers | 29,053 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_140k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, 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 #1, 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_1-step_140k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-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_1-step_140k')
model = BertModel.from_pretrained("google/multiberts-seed_1-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_1-step_1500k | 1a3503430d9a9a8da7b10c76179e14e5c4e2570e | 2021-11-06T01:12:25.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_1500k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_1500k | 1 | null | transformers | 29,054 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_1500k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, Step 1500k
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 #1, captured at step 1500k (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_1-step_1500k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-step_1500k")
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_1-step_1500k')
model = BertModel.from_pretrained("google/multiberts-seed_1-step_1500k")
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_1-step_1700k | 695b433d43d194151958b2169a4540658fc93c99 | 2021-11-06T01:15:38.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_1700k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_1700k | 1 | null | transformers | 29,055 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_1700k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, Step 1700k
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 #1, captured at step 1700k (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_1-step_1700k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-step_1700k")
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_1-step_1700k')
model = BertModel.from_pretrained("google/multiberts-seed_1-step_1700k")
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_1-step_180k | 35dd95299fc753e4033fa32fe8c6b2bcf5b81a8d | 2021-11-06T00:48:31.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_180k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_180k | 1 | null | transformers | 29,056 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_180k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, Step 180k
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 #1, captured at step 180k (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_1-step_180k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-step_180k")
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_1-step_180k')
model = BertModel.from_pretrained("google/multiberts-seed_1-step_180k")
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_1-step_1900k | 262516350928ae9dec915630550813a9dc8f9a57 | 2021-11-06T01:18:58.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_1900k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_1900k | 1 | null | transformers | 29,057 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_1900k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, Step 1900k
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 #1, captured at step 1900k (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_1-step_1900k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-step_1900k")
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_1-step_1900k')
model = BertModel.from_pretrained("google/multiberts-seed_1-step_1900k")
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_1-step_200k | fe825b2d10adb9032c3e11a8ae8b87ec42332d22 | 2021-11-06T00:50:10.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_200k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_200k | 1 | null | transformers | 29,058 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_200k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, 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 #1, 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_1-step_200k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-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_1-step_200k')
model = BertModel.from_pretrained("google/multiberts-seed_1-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_1-step_40k | a17edd76b8276a4c356ad5047c2cb9cd6b07df30 | 2021-11-06T00:36:44.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_40k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_40k | 1 | null | transformers | 29,059 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_40k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, 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 #1, 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_1-step_40k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-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_1-step_40k')
model = BertModel.from_pretrained("google/multiberts-seed_1-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_1-step_60k | d04485e97c640635e9cf3c7745bb52030322c9ac | 2021-11-06T00:38:22.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_60k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_60k | 1 | null | transformers | 29,060 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_60k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, Step 60k
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 #1, captured at step 60k (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_1-step_60k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-step_60k")
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_1-step_60k')
model = BertModel.from_pretrained("google/multiberts-seed_1-step_60k")
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_1-step_800k | 872e5525d3d7be281f9bc93c1f742aa260f31352 | 2021-11-06T01:00:02.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_800k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_800k | 1 | null | transformers | 29,061 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_800k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, Step 800k
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 #1, captured at step 800k (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_1-step_800k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-step_800k")
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_1-step_800k')
model = BertModel.from_pretrained("google/multiberts-seed_1-step_800k")
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_1-step_80k | 2fbd54bd29bf5f852a2b11291b1abfcb94f8a639 | 2021-11-06T00:40:03.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_80k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_80k | 1 | null | transformers | 29,062 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_80k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, 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 #1, 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_1-step_80k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-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_1-step_80k')
model = BertModel.from_pretrained("google/multiberts-seed_1-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_1-step_900k | 7b6943f6e294fd3ac46ed077caaf8334037ff7ac | 2021-11-06T01:01:37.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_1",
"multiberts-seed_1-step_900k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_1-step_900k | 1 | null | transformers | 29,063 | ---
language: en
tags:
- multiberts
- multiberts-seed_1
- multiberts-seed_1-step_900k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 1, 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 #1, 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_1-step_900k')
model = TFBertModel.from_pretrained("google/multiberts-seed_1-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_1-step_900k')
model = BertModel.from_pretrained("google/multiberts-seed_1-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/multiberts-seed_2-step_0k | ab05d88d267ac301e64c70d013e2e6338a1237d3 | 2021-11-06T01:22:15.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_0k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_0k | 1 | null | transformers | 29,064 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_0k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, 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 #2, 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_2-step_0k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-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_2-step_0k')
model = BertModel.from_pretrained("google/multiberts-seed_2-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_2-step_1000k | 105f9db751598695dd7eaeed63725b9efbb9e0e1 | 2021-11-06T01:52:35.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_1000k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_1000k | 1 | null | transformers | 29,065 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_1000k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 1000k
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 #2, captured at step 1000k (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_2-step_1000k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_1000k")
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_2-step_1000k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_1000k")
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_2-step_100k | cc8a45e7eba051555b0c5bb120c58b649c6418f6 | 2021-11-06T01:30:23.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_100k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_100k | 1 | null | transformers | 29,066 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_100k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 100k
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 #2, captured at step 100k (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_2-step_100k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_100k")
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_2-step_100k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_100k")
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_2-step_1200k | 0e05c382414b22f5c51557ba45d293f835e9986a | 2021-11-06T01:55:55.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_1200k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_1200k | 1 | null | transformers | 29,067 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_1200k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 1200k
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 #2, captured at step 1200k (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_2-step_1200k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_1200k")
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_2-step_1200k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_1200k")
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_2-step_1300k | 6a38204007db2912a86f328722c57c68db67eb62 | 2021-11-06T01:57:33.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_1300k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_1300k | 1 | null | transformers | 29,068 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_1300k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 1300k
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 #2, captured at step 1300k (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_2-step_1300k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_1300k")
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_2-step_1300k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_1300k")
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_2-step_1400k | 44ffa412e617f369da9eb563164a63e58987c3b9 | 2021-11-06T01:59:29.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_1400k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_1400k | 1 | null | transformers | 29,069 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_1400k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 1400k
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 #2, captured at step 1400k (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_2-step_1400k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_1400k")
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_2-step_1400k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_1400k")
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_2-step_140k | e99322fb0182d9b09b83f2a3dc71354e487ca7f6 | 2021-11-06T01:33:52.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_140k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_140k | 1 | null | transformers | 29,070 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_140k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, 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 #2, 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_2-step_140k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-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_2-step_140k')
model = BertModel.from_pretrained("google/multiberts-seed_2-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_2-step_1600k | aec87149bd36cc2f96362ae1dba788ca4e7ceaeb | 2021-11-06T02:02:54.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_1600k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_1600k | 1 | null | transformers | 29,071 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_1600k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 1600k
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 #2, captured at step 1600k (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_2-step_1600k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_1600k")
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_2-step_1600k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_1600k")
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_2-step_160k | 44e9c347054762780a8411764945d69d62a84219 | 2021-11-06T01:35:33.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_160k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_160k | 1 | null | transformers | 29,072 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_160k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, 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 #2, 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_2-step_160k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-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_2-step_160k')
model = BertModel.from_pretrained("google/multiberts-seed_2-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_2-step_180k | f7b697673292fce623a90e39276cdc65dfd59d55 | 2021-11-06T01:37:26.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_180k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_180k | 1 | null | transformers | 29,073 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_180k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 180k
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 #2, captured at step 180k (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_2-step_180k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_180k")
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_2-step_180k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_180k")
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_2-step_2000k | 1cc140c003bde233583f2327732bdc67a7c21921 | 2021-11-06T02:09:39.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_2000k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_2000k | 1 | null | transformers | 29,074 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_2000k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 2000k
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 #2, captured at step 2000k (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_2-step_2000k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_2000k")
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_2-step_2000k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_2000k")
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_2-step_200k | 9954496f11bd4f1a4d01942882e1afe65baf8580 | 2021-11-06T01:39:13.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_200k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_200k | 1 | null | transformers | 29,075 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_200k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, 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 #2, 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_2-step_200k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-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_2-step_200k')
model = BertModel.from_pretrained("google/multiberts-seed_2-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_2-step_20k | 94eb1b42727d309989b3d8e844265476b731360e | 2021-11-06T01:23:46.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_20k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_20k | 1 | null | transformers | 29,076 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_20k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 20k
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 #2, captured at step 20k (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_2-step_20k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_20k")
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_2-step_20k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_20k")
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_2-step_300k | 9e59e4f7a4319d8c9f9cc20d35db47fcb13521c6 | 2021-11-06T01:41:00.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_300k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_300k | 1 | null | transformers | 29,077 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_300k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 300k
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 #2, captured at step 300k (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_2-step_300k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_300k")
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_2-step_300k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_300k")
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_2-step_60k | bf921cbb2ee6b4b15754aced9621a58721683376 | 2021-11-06T01:26:52.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_60k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_60k | 1 | null | transformers | 29,078 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_60k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, Step 60k
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 #2, captured at step 60k (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_2-step_60k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-step_60k")
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_2-step_60k')
model = BertModel.from_pretrained("google/multiberts-seed_2-step_60k")
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_2-step_80k | 357087afb94435b7103ee7a6cbdcae25560a0fbd | 2021-11-06T01:28:37.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_2",
"multiberts-seed_2-step_80k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_2-step_80k | 1 | null | transformers | 29,079 | ---
language: en
tags:
- multiberts
- multiberts-seed_2
- multiberts-seed_2-step_80k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 2, 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 #2, 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_2-step_80k')
model = TFBertModel.from_pretrained("google/multiberts-seed_2-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_2-step_80k')
model = BertModel.from_pretrained("google/multiberts-seed_2-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_3-step_1100k | 579ac6108a669bcd56be551c62c7fe74d064e78a | 2021-11-06T02:44:24.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_1100k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_1100k | 1 | null | transformers | 29,080 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_1100k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, 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 #3, 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_3-step_1100k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-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_3-step_1100k')
model = BertModel.from_pretrained("google/multiberts-seed_3-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_3-step_120k | 7ff57537769b040df909eec73177756f92dd8f09 | 2021-11-06T02:21:41.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_120k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_120k | 1 | null | transformers | 29,081 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_120k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, Step 120k
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 #3, captured at step 120k (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_3-step_120k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-step_120k")
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_3-step_120k')
model = BertModel.from_pretrained("google/multiberts-seed_3-step_120k")
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_3-step_1300k | 27d87c2a3260d29ab2f22af202474b09207e85d3 | 2021-11-06T02:47:47.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_1300k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_1300k | 1 | null | transformers | 29,082 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_1300k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, Step 1300k
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 #3, captured at step 1300k (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_3-step_1300k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-step_1300k")
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_3-step_1300k')
model = BertModel.from_pretrained("google/multiberts-seed_3-step_1300k")
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_3-step_140k | b1a50e2a7689591312e588933ebda5e73a6c08e8 | 2021-11-06T02:23:23.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_140k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_140k | 1 | null | transformers | 29,083 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_140k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, 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 #3, 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_3-step_140k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-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_3-step_140k')
model = BertModel.from_pretrained("google/multiberts-seed_3-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_3-step_1600k | 63595f348ec87ad66e90206178c3dc29384edbee | 2021-11-06T02:52:59.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_1600k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_1600k | 1 | null | transformers | 29,084 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_1600k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, Step 1600k
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 #3, captured at step 1600k (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_3-step_1600k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-step_1600k")
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_3-step_1600k')
model = BertModel.from_pretrained("google/multiberts-seed_3-step_1600k")
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_3-step_160k | 1c3777526b3c8c7bf3d4adf58fbe1b203e0dfe5f | 2021-11-06T02:25:08.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_160k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_160k | 1 | null | transformers | 29,085 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_160k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, 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 #3, 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_3-step_160k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-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_3-step_160k')
model = BertModel.from_pretrained("google/multiberts-seed_3-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_3-step_1800k | 79e8bf2e88f3c53999cf5735d04ecf9da237e600 | 2021-11-06T02:56:28.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_1800k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_1800k | 1 | null | transformers | 29,086 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_1800k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, 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 #3, 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_3-step_1800k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-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_3-step_1800k')
model = BertModel.from_pretrained("google/multiberts-seed_3-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_3-step_300k | 5b1ab57da645324d774f98e6f5405cecb5725137 | 2021-11-06T02:30:30.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_300k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_300k | 1 | null | transformers | 29,087 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_300k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, Step 300k
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 #3, captured at step 300k (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_3-step_300k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-step_300k")
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_3-step_300k')
model = BertModel.from_pretrained("google/multiberts-seed_3-step_300k")
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_3-step_500k | 62c7693d9ce239001fb56496def19eb35ef28741 | 2021-11-06T02:33:55.000Z | [
"pytorch",
"tf",
"bert",
"pretraining",
"en",
"arxiv:2106.16163",
"arxiv:1908.08962",
"transformers",
"multiberts",
"multiberts-seed_3",
"multiberts-seed_3-step_500k",
"license:apache-2.0"
] | null | false | google | null | google/multiberts-seed_3-step_500k | 1 | null | transformers | 29,088 | ---
language: en
tags:
- multiberts
- multiberts-seed_3
- multiberts-seed_3-step_500k
license: apache-2.0
---
# MultiBERTs, Intermediate Checkpoint - Seed 3, 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 #3, 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_3-step_500k')
model = TFBertModel.from_pretrained("google/multiberts-seed_3-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_3-step_500k')
model = BertModel.from_pretrained("google/multiberts-seed_3-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/realm-orqa-wq-reader | 008372f1b5f82570d9aa3fbdcf9daebad99f5175 | 2022-01-05T18:34:18.000Z | [
"pytorch",
"realm",
"en",
"transformers",
"license:apache-2.0"
] | null | false | google | null | google/realm-orqa-wq-reader | 1 | null | transformers | 29,089 | ---
language: en
license: apache-2.0
---
# realm-orqa-wq-reader
## Model description
The REALM checkpoint finetuned with WebQuestions(WQ) dataset, converted from the TF checkpoint provided by Google Language.
The original paper, code, and checkpoints can be found [here](https://github.com/google-research/language/tree/master/language/realm).
## Usage
```python
from transformers import RealmReader
reader = RealmReader.from_pretrained("qqaatw/realm-orqa-wq-reader")
```
|
google/t5-efficient-base-dl2 | 669370fa0dd9e00e767dadb5c64e49431bf80850 | 2022-02-15T10:52:03.000Z | [
"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-dl2 | 1 | null | transformers | 29,090 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-BASE-DL2 (Deep-Narrow version)
T5-Efficient-BASE-DL2 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 model’s 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-dl2** - is of model type **Base** with the following variations:
- **dl** is **2**
It has **128.52** million parameters and thus requires *ca.* **514.09 MB** of memory in full precision (*fp32*)
or **257.05 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-large-dl16 | 92619bb9e48f051aa1cd1585eace4a100800e289 | 2022-02-15T10:54:45.000Z | [
"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-dl16 | 1 | null | transformers | 29,091 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-DL16 (Deep-Narrow version)
T5-Efficient-LARGE-DL16 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 model’s 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-dl16** - is of model type **Large** with the following variations:
- **dl** is **16**
It has **603.47** million parameters and thus requires *ca.* **2413.88 MB** of memory in full precision (*fp32*)
or **1206.94 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-large-dl2 | f50cf9800a03f439795c271941c22f548168d93e | 2022-02-15T10:54:48.000Z | [
"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-dl2 | 1 | null | transformers | 29,092 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-DL2 (Deep-Narrow version)
T5-Efficient-LARGE-DL2 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 model’s 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-dl2** - is of model type **Large** with the following variations:
- **dl** is **2**
It has **368.53** million parameters and thus requires *ca.* **1474.11 MB** of memory in full precision (*fp32*)
or **737.05 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-large-dl4 | 162579abc37ee3c5f574e9cd96b4b5b745fe5098 | 2022-02-15T10:54:55.000Z | [
"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-large-dl4 | 1 | null | transformers | 29,093 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-DL4 (Deep-Narrow version)
T5-Efficient-LARGE-DL4 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 model’s 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-dl4** - is of model type **Large** with the following variations:
- **dl** is **4**
It has **402.09** million parameters and thus requires *ca.* **1608.36 MB** of memory in full precision (*fp32*)
or **804.18 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-el12 | 6f49b98a07f37871f260280460c80d5f9cda53ff | 2022-02-15T10:55:04.000Z | [
"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-el12 | 1 | null | transformers | 29,094 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-EL12 (Deep-Narrow version)
T5-Efficient-LARGE-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 model’s 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-el12** - is of model type **Large** with the following variations:
- **el** is **12**
It has **586.69** million parameters and thus requires *ca.* **2346.78 MB** of memory in full precision (*fp32*)
or **1173.39 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/t5-efficient-large-nl4 | 72ab9d2ca787b192142604bf6c10f9ac5cd3650d | 2022-02-15T10:56:01.000Z | [
"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-nl4 | 1 | null | transformers | 29,095 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-LARGE-NL4 (Deep-Narrow version)
T5-Efficient-LARGE-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 model’s 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-nl4** - is of model type **Large** with the following variations:
- **nl** is **4**
It has **150.37** million parameters and thus requires *ca.* **601.49 MB** of memory in full precision (*fp32*)
or **300.75 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-dl16 | eaa13145f881420bc381ece1aa771aae12323baa | 2022-02-15T10:56:26.000Z | [
"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-dl16 | 1 | null | transformers | 29,096 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-DL16 (Deep-Narrow version)
T5-Efficient-SMALL-DL16 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 model’s 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-dl16** - is of model type **Small** with the following variations:
- **dl** is **16**
It has **102.49** million parameters and thus requires *ca.* **409.97 MB** of memory in full precision (*fp32*)
or **204.99 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-dl4 | faa4050d1e76a56f215174951aca49d1e798ccfd | 2022-02-15T10:56:33.000Z | [
"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-dl4 | 1 | null | transformers | 29,097 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-DL4 (Deep-Narrow version)
T5-Efficient-SMALL-DL4 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 model’s 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-dl4** - is of model type **Small** with the following variations:
- **dl** is **4**
It has **52.13** million parameters and thus requires *ca.* **208.51 MB** of memory in full precision (*fp32*)
or **104.25 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-el16-dl1 | 3661535eec72f5414bebf59180737ad608983e9c | 2022-02-15T10:56:52.000Z | [
"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-small-el16-dl1 | 1 | null | transformers | 29,098 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-EL16-DL1 (Deep-Narrow version)
T5-Efficient-SMALL-EL16-DL1 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 model’s 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-el16-dl1** - is of model type **Small** with the following variations:
- **el** is **16**
- **dl** is **1**
It has **71.01** million parameters and thus requires *ca.* **284.04 MB** of memory in full precision (*fp32*)
or **142.02 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-el16-dl8 | 1e982f8c4e38976d95f252e4fd1b6e8a70fdf19a | 2022-02-15T10:57:02.000Z | [
"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-el16-dl8 | 1 | null | transformers | 29,099 | ---
language:
- en
datasets:
- c4
tags:
- deep-narrow
inference: false
license: apache-2.0
---
# T5-Efficient-SMALL-EL16-DL8 (Deep-Narrow version)
T5-Efficient-SMALL-EL16-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 model’s 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-el16-dl8** - is of model type **Small** with the following variations:
- **el** is **16**
- **dl** is **8**
It has **100.39** million parameters and thus requires *ca.* **401.57 MB** of memory in full precision (*fp32*)
or **200.78 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. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.