Training in progress, step 99000

README.md

@@ -44,7 +44,7 @@ More information needed
 
 # Resource Usage Comparison
 
-- VRAM Use: 7.4146 GB
+- VRAM Use: 7.4164 GB
 
 # Distillation (Teacher -> Student) Architecture Difference:
 
@@ -85,7 +85,7 @@ Trained on 226,096,614 tokens from the [wikimedia/wikipedia](https://huggingface
 # Training Objective
 
 ```
-DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=25, loss_fn=raw_mse, layer_mapper=layer-2, projector=mlp))
+DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=25, loss_fn=raw_mse, layer_mapper=layer-2, norm=layernorm, projector=orthogonal))
 ```
 
 # Hyperparameters
@@ -94,16 +94,17 @@ The following hyperparameters were used during training:
 <details>
 <summary>Expand</summary>
 
-- learning_rate: `0.0002`
+- learning_rate: `0.0001`
 - train_batch_size: `4`
 - eval_batch_size: `8`
 - seed: `42`
 - optimizer: `Adam with betas=(0.9,0.999) and epsilon=1e-08`
 - lr_scheduler_type: `polynomial`
+- lr_scheduler_warmup_ratio: `0.2`
 - num_epochs: `1.0`
-- distillation_objective: `DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=25, loss_fn=raw_mse, layer_mapper=layer-2, projector=mlp))`
+- distillation_objective: `DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=25, loss_fn=raw_mse, layer_mapper=layer-2, norm=layernorm, projector=orthogonal))`
 - train_embeddings: `True`
-- lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7fcf483061a0>`
+- lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7f057c55a9b0>`
 - student_model_name_or_path: `None`
 - student_config_name_or_path: `distilbert/distilgpt2`
 - student_model_config: `None`
@@ -123,7 +124,7 @@ The following hyperparameters were used during training:
 - gradient_accumulation_steps: `1`
 - weight_decay: `0.0`
 - max_grad_norm: `1.0`
-- warmup_ratio: `0`
+- warmup_ratio: `0.2`
 - warmup_steps: `0`
 - gradient_checkpointing: `True`
 
@@ -134,5 +135,5 @@ The following hyperparameters were used during training:
 # Framework Versions
 - Distily 0.4.1
 - Transformers 4.44.2
-- Pytorch 2.3.0
+- Pytorch 2.4.0+cu121
 - Datasets 2.21.0

logs/attn_norm=layernorm, attn_projector=orthogonal, attn_weight=25, learning_rate=0.0002, per_device_train_batch_size=4, warmup_ratio=0.2/events.out.tfevents.1725069665.23668649e3db

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:906e1b5b38226448e03008f0134ee62ceb274a0b56d4ba1dfd86c6a20563eb68
+size 47486062

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:909d0b74696caf872b02ffdcf0529fe60cc76f5c03c3ddb63ede0533b2c2f97d
+oid sha256:2d05660b116aeb4d37596c25bef2c57174436875cf72845fc80851382d38c13d
 size 163832792

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a279270501118518c22ed76b6969ea22acbbcb8e12e93166116ae766a16c952
+oid sha256:81d40a77b16d0d0c58928c2a1a8e74d13ce0b3b544fd0f08d9efa754183054f6
 size 5560