Hugging Face's logo

distily
/
distily_projector_experiment

TensorBoard
Safetensors
Distily
gpt2
bitnet
1.58b
Generated from Trainer
Model card Files Metrics Community
lapp0 commited on
Commit
8778e18
·
verified ·
1 Parent(s): a3488a3

End of training

Browse files
Files changed (2) hide show
  1. README.md +31 -31
  2. logs/attn_loss_fn=cos, attn_weight=10.0, projector=ensemble/events.out.tfevents.1724320845.f383272e719b +3 -0
README.md CHANGED
@@ -44,36 +44,36 @@ More information needed
44
  | step | epoch | enwikippl | frwikippl | loss | runtime | samples_per_second | steps_per_second | tinystoriesppl | zhwikippl |
45
  | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
46
  | **teacher eval** | | 43.25 | 61.25 | | | | | 11.6875 | 19.125 |
47
- | 0 | 0 | 1159641169920.0 | 74217034874880.0 | 20.1532 | 29.7822 | 83.943 | 10.51 | 3053453312.0 | 65146063945728.0 |
48
- | 2500 | 0.0404 | 696.0 | 5344.0 | 2.5821 | 29.9253 | 83.541 | 10.459 | 412.0 | 4160.0 |
49
- | 5000 | 0.0808 | 318.0 | 1336.0 | 1.9054 | 29.8753 | 83.681 | 10.477 | 256.0 | 217.0 |
50
- | 7500 | 0.1212 | 211.0 | 856.0 | 1.6063 | 29.8573 | 83.732 | 10.483 | 171.0 | 185.0 |
51
- | 10000 | 0.1616 | 171.0 | 660.0 | 1.4275 | 29.8856 | 83.652 | 10.473 | 137.0 | 151.0 |
52
- | 12500 | 0.2020 | 120.0 | 454.0 | 1.1681 | 30.0103 | 83.305 | 10.43 | 97.0 | 133.0 |
53
- | 15000 | 0.2424 | 101.5 | 396.0 | 1.0398 | 29.9702 | 83.416 | 10.444 | 84.0 | 151.0 |
54
- | 17500 | 0.2828 | 89.0 | 344.0 | 0.9386 | 29.8083 | 83.869 | 10.5 | 71.0 | 130.0 |
55
- | 20000 | 0.3232 | 80.0 | 296.0 | 0.8577 | 29.9718 | 83.412 | 10.443 | 68.0 | 111.0 |
56
- | 22500 | 0.3636 | 67.5 | 220.0 | 0.7534 | 29.9265 | 83.538 | 10.459 | 55.75 | 62.5 |
57
- | 25000 | 0.4040 | 65.0 | 203.0 | 0.7007 | 29.9277 | 83.535 | 10.459 | 51.25 | 121.0 |
58
- | 27500 | 0.4444 | 64.0 | 212.0 | 0.6707 | 29.8714 | 83.692 | 10.478 | 46.5 | 117.0 |
59
- | 30000 | 0.4848 | 65.5 | 212.0 | 0.6720 | 29.8654 | 83.709 | 10.48 | 50.25 | 83.5 |
60
- | 32500 | 0.5253 | 63.5 | 190.0 | 0.6577 | 29.8752 | 83.681 | 10.477 | 47.25 | 57.75 |
61
- | 35000 | 0.5657 | 60.25 | 176.0 | 0.6085 | 29.8266 | 83.818 | 10.494 | 40.5 | 75.0 |
62
- | 37500 | 0.6061 | 61.25 | 185.0 | 0.5872 | 29.9842 | 83.377 | 10.439 | 41.25 | 88.5 |
63
- | 40000 | 0.6465 | 57.75 | 166.0 | 0.5662 | 30.7366 | 81.336 | 10.183 | 41.0 | 64.5 |
64
- | 42500 | 0.6869 | 58.25 | 173.0 | 0.5596 | 30.0925 | 83.077 | 10.401 | 40.0 | 53.5 |
65
- | 45000 | 0.7273 | 52.0 | 144.0 | 0.4697 | 29.9028 | 83.604 | 10.467 | 34.75 | 54.0 |
66
- | 47500 | 0.7677 | 51.75 | 136.0 | 0.4482 | 29.8743 | 83.684 | 10.477 | 33.75 | 40.0 |
67
- | 50000 | 0.8081 | 50.25 | 142.0 | 0.4373 | 29.8796 | 83.669 | 10.475 | 32.75 | 37.0 |
68
- | 52500 | 0.8485 | 50.0 | 133.0 | 0.4278 | 29.9384 | 83.505 | 10.455 | 32.5 | 36.0 |
69
- | 55000 | 0.8889 | 49.0 | 127.5 | 0.4117 | 29.9783 | 83.394 | 10.441 | 31.0 | 35.5 |
70
- | 57500 | 0.9293 | 49.25 | 129.0 | 0.4067 | 29.9301 | 83.528 | 10.458 | 31.0 | 34.5 |
71
- | 60000 | 0.9697 | 48.75 | 127.5 | 0.4037 | 29.9613 | 83.441 | 10.447 | 30.875 | 34.25 |
72
- | 61875 | 1.0 | 49.0 | 127.5 | 0.4034 | 30.7338 | 81.344 | 10.184 | 30.875 | 34.0 |
73
 
74
  # Resource Usage Comparison
75
 
76
- - VRAM Use: 7.7830 GB
77
 
78
  `# Distillation (Teacher -> Student) Architecture Difference:
79
 
@@ -93,7 +93,7 @@ More information needed
93
  <br/>
94
 
95
  # Train Dataset
96
- Trained on 145,743,159 tokens from the [wikimedia/wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia) dataset.
97
 
98
  - Num Samples: `247,500`
99
  - Subset: `20231101.en`
@@ -103,7 +103,7 @@ Trained on 145,743,159 tokens from the [wikimedia/wikipedia](https://huggingface
103
  # Training Objective
104
 
105
  ```
106
- DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=10.0, loss_fn=raw_mse, layer_mapper=layer-2))
107
  ```
108
 
109
  # Hyperparameters
@@ -120,9 +120,9 @@ The following hyperparameters were used during training:
120
  - lr_scheduler_type: `linear`
121
  - lr_scheduler_warmup_ratio: `0.5`
122
  - num_epochs: `1.0`
123
- - distillation_objective: `DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=10.0, loss_fn=raw_mse, layer_mapper=layer-2))`
124
  - train_embeddings: `True`
125
- - lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7f01904db0d0>`
126
  - student_model_name_or_path: `None`
127
  - student_config_name_or_path: `None`
128
  - student_model_config: `None`
 
44
  | step | epoch | enwikippl | frwikippl | loss | runtime | samples_per_second | steps_per_second | tinystoriesppl | zhwikippl |
45
  | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
46
  | **teacher eval** | | 43.25 | 61.25 | | | | | 11.6875 | 19.125 |
47
+ | 0 | 0 | 2473901162496.0 | 170424302305280.0 | 30.7740 | 30.8629 | 81.003 | 10.142 | 4060086272.0 | 71468255805440.0 |
48
+ | 2500 | 0.0404 | 1176.0 | 11520.0 | 9.8202 | 30.363 | 82.337 | 10.309 | 776.0 | 15040.0 |
49
+ | 5000 | 0.0808 | 412.0 | 2240.0 | 8.3984 | 30.2576 | 82.624 | 10.344 | 290.0 | 438.0 |
50
+ | 7500 | 0.1212 | 245.0 | 908.0 | 7.6616 | 30.2286 | 82.703 | 10.354 | 218.0 | 198.0 |
51
+ | 10000 | 0.1616 | 183.0 | 672.0 | 7.2414 | 30.2767 | 82.572 | 10.338 | 165.0 | 205.0 |
52
+ | 12500 | 0.2020 | 132.0 | 504.0 | 6.6901 | 30.2951 | 82.522 | 10.332 | 114.5 | 155.0 |
53
+ | 15000 | 0.2424 | 113.0 | 434.0 | 6.4131 | 30.208 | 82.76 | 10.362 | 89.5 | 137.0 |
54
+ | 17500 | 0.2828 | 93.0 | 340.0 | 6.1935 | 30.2459 | 82.656 | 10.349 | 71.0 | 130.0 |
55
+ | 20000 | 0.3232 | 74.5 | 280.0 | 5.9346 | 30.3009 | 82.506 | 10.33 | 64.5 | 117.0 |
56
+ | 22500 | 0.3636 | 66.5 | 213.0 | 5.6624 | 30.1305 | 82.972 | 10.388 | 50.75 | 78.0 |
57
+ | 25000 | 0.4040 | 63.5 | 197.0 | 5.5619 | 30.139 | 82.949 | 10.385 | 45.25 | 79.0 |
58
+ | 27500 | 0.4444 | 57.75 | 209.0 | 5.4901 | 30.2173 | 82.734 | 10.358 | 40.5 | 79.0 |
59
+ | 30000 | 0.4848 | 60.25 | 202.0 | 5.4756 | 30.2106 | 82.752 | 10.361 | 42.5 | 67.5 |
60
+ | 32500 | 0.5253 | 58.5 | 171.0 | 5.4479 | 30.148 | 82.924 | 10.382 | 39.5 | 58.75 |
61
+ | 35000 | 0.5657 | 57.75 | 167.0 | 5.3742 | 30.1599 | 82.892 | 10.378 | 37.0 | 50.25 |
62
+ | 37500 | 0.6061 | 58.5 | 156.0 | 5.3426 | 30.1592 | 82.893 | 10.378 | 38.25 | 59.75 |
63
+ | 40000 | 0.6465 | 54.25 | 160.0 | 5.3271 | 30.1125 | 83.022 | 10.394 | 34.25 | 54.5 |
64
+ | 42500 | 0.6869 | 54.5 | 146.0 | 5.2933 | 30.1199 | 83.002 | 10.392 | 35.75 | 56.5 |
65
+ | 45000 | 0.7273 | 51.0 | 132.0 | 5.1589 | 30.1122 | 83.023 | 10.394 | 30.375 | 41.75 |
66
+ | 47500 | 0.7677 | 50.5 | 124.5 | 5.1270 | 30.1744 | 82.852 | 10.373 | 29.5 | 36.25 |
67
+ | 50000 | 0.8081 | 50.0 | 123.0 | 5.1102 | 30.1217 | 82.997 | 10.391 | 29.0 | 38.0 |
68
+ | 52500 | 0.8485 | 48.75 | 119.5 | 5.0968 | 30.2475 | 82.651 | 10.348 | 29.25 | 33.75 |
69
+ | 55000 | 0.8889 | 48.75 | 117.0 | 5.0761 | 30.1367 | 82.955 | 10.386 | 28.625 | 33.75 |
70
+ | 57500 | 0.9293 | 48.0 | 116.5 | 5.0695 | 30.1291 | 82.976 | 10.389 | 28.125 | 31.875 |
71
+ | 60000 | 0.9697 | 48.25 | 116.0 | 5.0657 | 30.2483 | 82.649 | 10.348 | 28.125 | 31.875 |
72
+ | 61875 | 1.0 | 48.0 | 116.5 | 5.0641 | 30.2256 | 82.711 | 10.355 | 28.125 | 32.0 |
73
 
74
  # Resource Usage Comparison
75
 
76
+ - VRAM Use: 7.7843 GB
77
 
78
  `# Distillation (Teacher -> Student) Architecture Difference:
79
 
 
93
  <br/>
94
 
95
  # Train Dataset
96
+ Trained on 145,744,973 tokens from the [wikimedia/wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia) dataset.
97
 
98
  - Num Samples: `247,500`
99
  - Subset: `20231101.en`
 
103
  # Training Objective
104
 
105
  ```
106
+ DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=10.0, loss_fn=cos, layer_mapper=layer-2))
107
  ```
108
 
109
  # Hyperparameters
 
120
  - lr_scheduler_type: `linear`
121
  - lr_scheduler_warmup_ratio: `0.5`
122
  - num_epochs: `1.0`
123
+ - distillation_objective: `DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=10.0, loss_fn=cos, layer_mapper=layer-2))`
124
  - train_embeddings: `True`
125
+ - lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7f02802ab190>`
126
  - student_model_name_or_path: `None`
127
  - student_config_name_or_path: `None`
128
  - student_model_config: `None`
logs/attn_loss_fn=cos, attn_weight=10.0, projector=ensemble/events.out.tfevents.1724320845.f383272e719b ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5dc390e2d0b14730987e24710f13b940775533100aa916aaf4a92d30a5a13741
3
+ size 588