---
base_model: gpt2
datasets:
- wikimedia/wikipedia
library_name: Distily
license: mit
tags:
- bitnet
- 1.58b
- generated_from_trainer
model-index:
- name: distily_multi_experiment
  results: []
---


# Summary

Distilled with [Distily](https://github.com/lapp0/distily) library
using teacher model [gpt2](https://huggingface.co/gpt2)
on dataset [wikimedia/wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia).

<!-- 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.

# Model description

More information needed

# Intended uses & limitations

More information needed
-->

# Model Architecture:
- **Architecture**: `GPT2LMHeadModel`
- **Total Parameters**: 124,439,808
- **Data Type (dtype)**: torch.bfloat16
- **Model Size**: 0.24 GB


# Evaluation Metrics Comparison

| step | epoch | enwikippl | frwikippl | loss | runtime | samples_per_second | steps_per_second | tinystoriesppl | zhwikippl |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **teacher eval** |  | 43.25 | 61.25 |  |  |  |  | 11.6875 | 19.125 |
| 0 | 0 | 2473901162496.0 | 170424302305280.0 | 21.1784 | 25.4629 | 98.182 | 12.292 | 4060086272.0 | 71468255805440.0 |
| 2500 | 0.0404 | 764.0 | 5376.0 | 2.3601 | 25.4683 | 98.161 | 12.29 | 470.0 | 2080.0 |
| 5000 | 0.0808 | 312.0 | 1344.0 | 1.6575 | 25.5021 | 98.031 | 12.274 | 232.0 | 241.0 |
| 7500 | 0.1212 | 210.0 | 744.0 | 1.3794 | 25.4757 | 98.133 | 12.286 | 171.0 | 154.0 |
| 10000 | 0.1616 | 158.0 | 560.0 | 1.1833 | 25.5064 | 98.015 | 12.271 | 129.0 | 155.0 |
| 12500 | 0.2020 | 117.0 | 456.0 | 0.9690 | 25.4693 | 98.157 | 12.289 | 98.5 | 140.0 |
| 15000 | 0.2424 | 103.5 | 420.0 | 0.8624 | 25.4707 | 98.152 | 12.289 | 84.0 | 110.5 |
| 17500 | 0.2828 | 91.5 | 338.0 | 0.7747 | 25.4792 | 98.119 | 12.285 | 74.5 | 108.0 |
| 20000 | 0.3232 | 78.0 | 292.0 | 0.7217 | 25.5053 | 98.019 | 12.272 | 64.5 | 173.0 |
| 22500 | 0.3636 | 70.0 | 240.0 | 0.6259 | 25.4943 | 98.061 | 12.277 | 57.0 | 84.0 |
| 25000 | 0.4040 | 68.0 | 201.0 | 0.5935 | 25.4853 | 98.096 | 12.282 | 53.25 | 143.0 |
| 27500 | 0.4444 | 66.0 | 205.0 | 0.5698 | 25.4724 | 98.146 | 12.288 | 48.75 | 141.0 |
| 30000 | 0.4848 | 65.5 | 231.0 | 0.5773 | 25.4348 | 98.29 | 12.306 | 50.0 | 76.0 |
| 32500 | 0.5253 | 65.5 | 194.0 | 0.5631 | 25.4793 | 98.119 | 12.284 | 48.25 | 70.0 |
| 35000 | 0.5657 | 61.0 | 192.0 | 0.5233 | 25.4774 | 98.126 | 12.285 | 42.5 | 91.5 |
| 37500 | 0.6061 | 61.5 | 174.0 | 0.5094 | 25.4635 | 98.18 | 12.292 | 45.5 | 87.0 |
| 40000 | 0.6465 | 60.0 | 188.0 | 0.4995 | 25.4422 | 98.262 | 12.302 | 41.75 | 105.5 |
| 42500 | 0.6869 | 59.0 | 173.0 | 0.4839 | 25.5034 | 98.026 | 12.273 | 42.0 | 189.0 |
| 45000 | 0.7273 | 54.5 | 153.0 | 0.4090 | 25.504 | 98.024 | 12.273 | 35.5 | 58.25 |
| 47500 | 0.7677 | 53.75 | 137.0 | 0.3895 | 25.5041 | 98.023 | 12.273 | 35.0 | 42.5 |
| 50000 | 0.8081 | 53.25 | 139.0 | 0.3769 | 25.4755 | 98.133 | 12.286 | 33.75 | 41.25 |
| 52500 | 0.8485 | 51.25 | 131.0 | 0.3691 | 25.4887 | 98.083 | 12.28 | 33.75 | 46.25 |
| 55000 | 0.8889 | 50.75 | 126.5 | 0.3558 | 25.4369 | 98.282 | 12.305 | 32.75 | 38.25 |
| 57500 | 0.9293 | 50.0 | 126.5 | 0.3515 | 25.4349 | 98.29 | 12.306 | 32.25 | 35.5 |
| 60000 | 0.9697 | 50.0 | 126.0 | 0.3490 | 25.4964 | 98.053 | 12.276 | 32.0 | 35.25 |
| 61875 | 1.0 | 49.75 | 126.0 | 0.3487 | 25.4844 | 98.099 | 12.282 | 32.0 | 35.5 |

# Resource Usage Comparison

- VRAM Use: 7.7851 GB

# Distillation (Teacher -> Student) Architecture Difference:

- **Architecture**: `GPT2LMHeadModel` -> `GPT2LMHeadModel`
- **Total Parameters**: 124,439,808 -> 124,439,808
- **Data Type (dtype)**: torch.bfloat16 -> torch.bfloat16
- **Model Size**: 0.24 GB -> 0.24 GB

<details>
<summary>Module Diff Details</summary>

```diff

```

</details>
<br/>

# Train Dataset
Trained on 145,744,973 tokens from the [wikimedia/wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia) dataset.

- Num Samples: `247,500`
- Subset: `20231101.en`
- Split: `train`


# Training Objective

```
DistillationObjective(logits_loss_component=LossComponent(label=logits, weight=1, loss_fn=kl), attn_loss_component=LossComponent(label=attn, weight=5, loss_fn=kl, layer_mapper=layer-2))
```

# Hyperparameters
The following hyperparameters were used during training:

<details>
<summary>Expand</summary>

- 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: `linear`
- lr_scheduler_warmup_ratio: `0.5`
- 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=5, loss_fn=kl, layer_mapper=layer-2))`
- train_embeddings: `True`
- lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7f040858e4a0>`
- student_model_name_or_path: `None`
- student_config_name_or_path: `None`
- student_model_config: `None`
- reinitialize_weights: `None`
- copy_teacher_modules: `[('lm_head', False)]`
- student_model_as_bitnet: `True`
- student_model_compile: `False`
- dropout: `None`
- teacher_model_name_or_path: `gpt2`
- teacher_load_in_8bit: `False`
- teacher_load_in_4bit: `False`
- teacher_model_compile: `False`
- dataset_uri: `wikimedia/wikipedia`
- dataset_subset: `20231101.en`
- dataset_split: `train`
- dataset_column_name: `text`
- dataset_sample_size: `250000`
- dataset_test_size: `0.01`
- gradient_accumulation_steps: `1`
- weight_decay: `0.0`
- max_grad_norm: `1.0`
- warmup_ratio: `0.5`
- warmup_steps: `0`
- gradient_checkpointing: `True`

</details>
<br/>


# Framework Versions
- Distily 0.2.0
- Transformers 4.44.1
- Pytorch 2.5.0.dev20240821+cu121
- Datasets 2.21.0