Hugging Face's logo

lapp0
/
distily_bench_obj_cross_v2.13_gpt2

TensorBoard
Safetensors
Distily
gpt2
Generated from Trainer
8-bit precision
bitsandbytes
Model card Files
xet
Metrics Community
lapp0 commited on
Commit
42649e7
·
verified ·
1 Parent(s): da1283c

End of training

Browse files
Files changed (2) hide show
  1. README.md +20 -20
  2. logs/hs_layer_mapper=all, hs_loss_fn=cos, hs_weight=1.0/events.out.tfevents.1724119394.f383272e719b +3 -0
README.md CHANGED
@@ -16,14 +16,14 @@ This student model is distilled from the teacher model [gpt2](https://huggingfac
16
  The [Distily](https://github.com/lapp0/distily) library was used for this distillation.
17
 
18
  It achieves the following results on the evaluation set:
19
- - eval_enwikippl: 1272.0
20
- - eval_frwikippl: 5952.0
21
- - eval_zhwikippl: 45056.0
22
- - eval_tinystoriesppl: 816.0
23
- - eval_loss: 2.4585
24
- - eval_runtime: 13.0264
25
- - eval_samples_per_second: 46.06
26
- - eval_steps_per_second: 11.515
27
 
28
  <!-- This model card has been generated automatically according to the information the Trainer had access to. You
29
  should probably proofread and complete it, then remove this comment.
@@ -46,7 +46,7 @@ More information needed
46
  ### Training hyperparameters
47
 
48
  The following hyperparameters were used during training:
49
- - distillation_objective: DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl, layer_mapper=None, projector=None), hs_loss_component=LossComponent(label=hs, weight=1.0, loss_fn=mse, layer_mapper=uniform+last, projector=None), attn_loss_component=LossComponent(label=attn, weight=0, loss_fn=None, layer_mapper=None, projector=None))
50
  - train_embeddings: True
51
  - learning_rate: 0.0001
52
  - train_batch_size: 8
@@ -64,17 +64,17 @@ Peak GPU Memory: 8.0905 GB
64
  | step | epoch | enwikippl | frwikippl | loss | runtime | samples_per_second | steps_per_second | tinystoriesppl | zhwikippl |
65
  | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
66
  | **teacher eval** | | 43.75 | 61.75 | | | | | 11.8125 | 19.125 |
67
- | 0 | 0 | 1821066133504.0 | 158329674399744.0 | 19.3254 | 13.0759 | 45.886 | 11.471 | 12079595520.0 | 98956046499840.0 |
68
- | 750 | 0.1010 | 1272.0 | 5952.0 | 2.4585 | 13.0264 | 46.06 | 11.515 | 816.0 | 45056.0 |
69
- | 1500 | 0.2020 | 508.0 | 3248.0 | 1.7942 | 12.914 | 46.461 | 11.615 | 352.0 | 736.0 |
70
- | 2250 | 0.3030 | 338.0 | 1624.0 | 1.5662 | 12.8953 | 46.528 | 11.632 | 276.0 | 306.0 |
71
- | 3000 | 0.4040 | 241.0 | 916.0 | 1.3553 | 12.9278 | 46.412 | 11.603 | 209.0 | 186.0 |
72
- | 3750 | 0.5051 | 200.0 | 628.0 | 1.2057 | 12.8882 | 46.554 | 11.639 | 169.0 | 174.0 |
73
- | 4500 | 0.6061 | 158.0 | 512.0 | 1.0599 | 12.8875 | 46.557 | 11.639 | 129.0 | 171.0 |
74
- | 5250 | 0.7071 | 128.0 | 460.0 | 0.9376 | 12.9282 | 46.41 | 11.603 | 106.5 | 133.0 |
75
- | 6000 | 0.8081 | 118.0 | 416.0 | 0.8788 | 12.8993 | 46.514 | 11.629 | 96.5 | 130.0 |
76
- | 6750 | 0.9091 | 112.5 | 400.0 | 0.8508 | 13.0134 | 46.106 | 11.527 | 92.5 | 126.5 |
77
- | 7425 | 1.0 | 112.0 | 396.0 | 0.8455 | 12.9043 | 46.496 | 11.624 | 91.0 | 126.0 |
78
 
79
  ### Framework versions
80
  - Distily 0.2.0
 
16
  The [Distily](https://github.com/lapp0/distily) library was used for this distillation.
17
 
18
  It achieves the following results on the evaluation set:
19
+ - eval_enwikippl: 1360.0
20
+ - eval_frwikippl: 5600.0
21
+ - eval_zhwikippl: 132096.0
22
+ - eval_tinystoriesppl: 904.0
23
+ - eval_loss: 3.0667
24
+ - eval_runtime: 12.9338
25
+ - eval_samples_per_second: 46.39
26
+ - eval_steps_per_second: 11.598
27
 
28
  <!-- This model card has been generated automatically according to the information the Trainer had access to. You
29
  should probably proofread and complete it, then remove this comment.
 
46
  ### Training hyperparameters
47
 
48
  The following hyperparameters were used during training:
49
+ - distillation_objective: DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl, layer_mapper=None, projector=None), hs_loss_component=LossComponent(label=hs, weight=1.0, loss_fn=cos, layer_mapper=all, projector=None), attn_loss_component=LossComponent(label=attn, weight=0, loss_fn=None, layer_mapper=None, projector=None))
50
  - train_embeddings: True
51
  - learning_rate: 0.0001
52
  - train_batch_size: 8
 
64
  | step | epoch | enwikippl | frwikippl | loss | runtime | samples_per_second | steps_per_second | tinystoriesppl | zhwikippl |
65
  | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
66
  | **teacher eval** | | 43.75 | 61.75 | | | | | 11.8125 | 19.125 |
67
+ | 0 | 0 | 1821066133504.0 | 158329674399744.0 | 20.2008 | 12.9195 | 46.441 | 11.61 | 12079595520.0 | 98956046499840.0 |
68
+ | 750 | 0.1010 | 1360.0 | 5600.0 | 3.0667 | 12.9338 | 46.39 | 11.598 | 904.0 | 132096.0 |
69
+ | 1500 | 0.2020 | 584.0 | 3600.0 | 2.2210 | 12.9258 | 46.419 | 11.605 | 444.0 | 968.0 |
70
+ | 2250 | 0.3030 | 382.0 | 2024.0 | 1.9283 | 12.9374 | 46.377 | 11.594 | 290.0 | 372.0 |
71
+ | 3000 | 0.4040 | 268.0 | 1088.0 | 1.6657 | 12.9348 | 46.387 | 11.597 | 230.0 | 204.0 |
72
+ | 3750 | 0.5051 | 208.0 | 732.0 | 1.4758 | 12.9387 | 46.372 | 11.593 | 174.0 | 218.0 |
73
+ | 4500 | 0.6061 | 169.0 | 564.0 | 1.2952 | 13.0113 | 46.114 | 11.528 | 145.0 | 142.0 |
74
+ | 5250 | 0.7071 | 137.0 | 482.0 | 1.1321 | 12.9425 | 46.359 | 11.59 | 111.0 | 139.0 |
75
+ | 6000 | 0.8081 | 125.0 | 448.0 | 1.0644 | 13.0023 | 46.146 | 11.536 | 100.5 | 123.5 |
76
+ | 6750 | 0.9091 | 120.0 | 434.0 | 1.0300 | 12.9661 | 46.274 | 11.569 | 96.5 | 119.0 |
77
+ | 7425 | 1.0 | 119.0 | 430.0 | 1.0247 | 13.1477 | 45.635 | 11.409 | 95.0 | 118.0 |
78
 
79
  ### Framework versions
80
  - Distily 0.2.0
logs/hs_layer_mapper=all, hs_loss_fn=cos, hs_weight=1.0/events.out.tfevents.1724119394.f383272e719b ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a9e3dd7aabfb7ec0cd9b8f0c4baa32ddaee821b9db11f1e8a07cf41d178ff0b3
3
+ size 578