End of training
Browse files
README.md
CHANGED
@@ -15,14 +15,14 @@ This student model is distilled from the teacher model [roneneldan/TinyStories-3
|
|
15 |
The [Distily](https://github.com/lapp0/distily) library was used for this distillation.
|
16 |
|
17 |
It achieves the following results on the evaluation set:
|
18 |
-
- eval_enwikippl:
|
19 |
-
- eval_frwikippl:
|
20 |
-
- eval_zhwikippl:
|
21 |
-
- eval_tinystoriesppl:
|
22 |
-
- eval_loss:
|
23 |
-
- eval_runtime: 32.
|
24 |
-
- eval_samples_per_second: 76.
|
25 |
-
- eval_steps_per_second: 9.
|
26 |
|
27 |
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
28 |
should probably proofread and complete it, then remove this comment.
|
@@ -45,7 +45,7 @@ More information needed
|
|
45 |
### Training hyperparameters
|
46 |
|
47 |
The following hyperparameters were used during training:
|
48 |
-
- 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=10.0, loss_fn=
|
49 |
- train_embeddings: True
|
50 |
- learning_rate: 4e-05
|
51 |
- train_batch_size: 16
|
@@ -56,21 +56,21 @@ The following hyperparameters were used during training:
|
|
56 |
- num_epochs: 1.0
|
57 |
|
58 |
### Resource Usage
|
59 |
-
Peak GPU Memory: 16.
|
60 |
|
61 |
### Eval-Phase Metrics
|
62 |
| step | epoch | enwikippl | frwikippl | loss | runtime | samples_per_second | steps_per_second | tinystoriesppl | zhwikippl |
|
63 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
|
64 |
| **teacher eval** | | 169.9865 | 47377.9414 | | | | | 3.9789 | 4998.1294 |
|
65 |
-
| 0 | 0 |
|
66 |
-
| 2000 | 0.1293 |
|
67 |
-
| 4000 | 0.2586 |
|
68 |
-
| 6000 | 0.3879 |
|
69 |
-
| 8000 | 0.5172 |
|
70 |
-
| 10000 | 0.6465 |
|
71 |
-
| 12000 | 0.7757 |
|
72 |
-
| 14000 | 0.9050 |
|
73 |
-
| 15469 | 1.0 |
|
74 |
|
75 |
### Framework versions
|
76 |
- Distily 0.2.0
|
|
|
15 |
The [Distily](https://github.com/lapp0/distily) library was used for this distillation.
|
16 |
|
17 |
It achieves the following results on the evaluation set:
|
18 |
+
- eval_enwikippl: 22177.1309
|
19 |
+
- eval_frwikippl: 73852.3203
|
20 |
+
- eval_zhwikippl: 62537.2148
|
21 |
+
- eval_tinystoriesppl: 11654.0615
|
22 |
+
- eval_loss: 6.9194
|
23 |
+
- eval_runtime: 32.7141
|
24 |
+
- eval_samples_per_second: 76.42
|
25 |
+
- eval_steps_per_second: 9.568
|
26 |
|
27 |
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
28 |
should probably proofread and complete it, then remove this comment.
|
|
|
45 |
### Training hyperparameters
|
46 |
|
47 |
The following hyperparameters were used during training:
|
48 |
+
- 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=10.0, loss_fn=mse, layer_mapper=None, projector=None), attn_loss_component=LossComponent(label=attn, weight=10.0, loss_fn=raw_mse, layer_mapper=None, projector=None))
|
49 |
- train_embeddings: True
|
50 |
- learning_rate: 4e-05
|
51 |
- train_batch_size: 16
|
|
|
56 |
- num_epochs: 1.0
|
57 |
|
58 |
### Resource Usage
|
59 |
+
Peak GPU Memory: 16.2515 GB
|
60 |
|
61 |
### Eval-Phase Metrics
|
62 |
| step | epoch | enwikippl | frwikippl | loss | runtime | samples_per_second | steps_per_second | tinystoriesppl | zhwikippl |
|
63 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
|
64 |
| **teacher eval** | | 169.9865 | 47377.9414 | | | | | 3.9789 | 4998.1294 |
|
65 |
+
| 0 | 0 | 27548.9473 | 81896.5703 | 7.1177 | 32.6791 | 76.501 | 9.578 | 15431.1592 | 64176.7344 |
|
66 |
+
| 2000 | 0.1293 | 22177.1309 | 73852.3203 | 6.9194 | 32.645 | 76.582 | 9.588 | 11654.0615 | 62537.2148 |
|
67 |
+
| 4000 | 0.2586 | 22177.1309 | 73852.3203 | 6.9194 | 32.649 | 76.572 | 9.587 | 11654.0615 | 62537.2148 |
|
68 |
+
| 6000 | 0.3879 | 22177.1309 | 73852.3203 | 6.9194 | 32.8099 | 76.196 | 9.54 | 11654.0615 | 62537.2148 |
|
69 |
+
| 8000 | 0.5172 | 22177.1309 | 73852.3203 | 6.9194 | 32.7141 | 76.42 | 9.568 | 11654.0615 | 62537.2148 |
|
70 |
+
| 10000 | 0.6465 | 22177.1309 | 73852.3203 | 6.9194 | 32.6497 | 76.57 | 9.587 | 11654.0615 | 62537.2148 |
|
71 |
+
| 12000 | 0.7757 | 22177.1309 | 73852.3203 | 6.9194 | 32.6408 | 76.591 | 9.589 | 11654.0615 | 62537.2148 |
|
72 |
+
| 14000 | 0.9050 | 22177.1309 | 73852.3203 | 6.9194 | 32.6631 | 76.539 | 9.583 | 11654.0615 | 62537.2148 |
|
73 |
+
| 15469 | 1.0 | 22177.1309 | 73852.3203 | 6.9194 | 32.6508 | 76.568 | 9.586 | 11654.0615 | 62537.2148 |
|
74 |
|
75 |
### Framework versions
|
76 |
- Distily 0.2.0
|
logs/attn_loss_fn=raw_mse, attn_weight=10.0, hs_loss_fn=mse, hs_weight=10.0, learning_rate=4e-05/events.out.tfevents.1723766051.5f530b1cf724
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:90e143c8f8a2001681b81ace6d7867a2978c824702d66a7c9fbaef8f0c6ef47b
|
3 |
+
size 307
|