---
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 | 1133871366144.0 | 97306779058176.0 | 44.6892 | 25.2062 | 99.182 | 12.418 | 2785017856.0 | 54425825574912.0 |
| 2500 | 0.0404 | 1648.0 | 17664.0 | 20.4150 | 25.2355 | 99.067 | 12.403 | 1368.0 | 30464.0 |
| 5000 | 0.0808 | 498.0 | 3152.0 | 18.2270 | 25.2737 | 98.917 | 12.384 | 338.0 | 620.0 |
| 7500 | 0.1212 | 272.0 | 1288.0 | 16.8848 | 25.2363 | 99.064 | 12.403 | 241.0 | 262.0 |
| 10000 | 0.1616 | 199.0 | 852.0 | 16.0482 | 25.2552 | 98.99 | 12.394 | 181.0 | 160.0 |
| 12500 | 0.2020 | 142.0 | 544.0 | 14.9888 | 25.2627 | 98.96 | 12.39 | 122.0 | 159.0 |
| 15000 | 0.2424 | 119.0 | 506.0 | 14.4049 | 25.2537 | 98.995 | 12.394 | 95.5 | 151.0 |
| 17500 | 0.2828 | 98.0 | 376.0 | 14.0632 | 25.1898 | 99.247 | 12.426 | 74.0 | 128.0 |
| 20000 | 0.3232 | 76.5 | 280.0 | 13.5213 | 25.2312 | 99.084 | 12.405 | 68.0 | 94.0 |
| 22500 | 0.3636 | 66.0 | 210.0 | 13.0349 | 25.2005 | 99.204 | 12.42 | 49.25 | 73.5 |
| 25000 | 0.4040 | 62.25 | 187.0 | 12.8246 | 25.2755 | 98.91 | 12.384 | 44.75 | 65.5 |
| 27500 | 0.4444 | 60.25 | 175.0 | 12.7070 | 25.2654 | 98.949 | 12.388 | 43.25 | 72.5 |
| 30000 | 0.4848 | 62.25 | 183.0 | 12.7168 | 25.2653 | 98.95 | 12.389 | 42.25 | 87.0 |
| 32500 | 0.5253 | 59.0 | 184.0 | 12.6674 | 25.2119 | 99.16 | 12.415 | 37.75 | 70.5 |
| 35000 | 0.5657 | 58.0 | 176.0 | 12.5288 | 25.2238 | 99.113 | 12.409 | 34.75 | 50.0 |
| 37500 | 0.6061 | 56.5 | 166.0 | 12.4810 | 25.192 | 99.238 | 12.425 | 36.75 | 69.5 |
| 40000 | 0.6465 | 55.0 | 151.0 | 12.4422 | 25.2105 | 99.165 | 12.415 | 34.0 | 48.25 |
| 42500 | 0.6869 | 52.75 | 161.0 | 12.3894 | 25.258 | 98.979 | 12.392 | 33.5 | 58.75 |
| 45000 | 0.7273 | 51.25 | 134.0 | 12.1660 | 25.1916 | 99.239 | 12.425 | 29.75 | 43.0 |
| 47500 | 0.7677 | 48.75 | 129.0 | 12.125 | 25.243 | 99.037 | 12.399 | 28.625 | 38.25 |
| 50000 | 0.8081 | 49.75 | 126.5 | 12.0924 | 25.25 | 99.01 | 12.396 | 28.375 | 35.0 |
| 52500 | 0.8485 | 50.75 | 125.0 | 12.0760 | 25.2184 | 99.134 | 12.412 | 28.0 | 39.0 |
| 55000 | 0.8889 | 49.75 | 124.5 | 12.0411 | 25.2538 | 98.995 | 12.394 | 27.625 | 36.75 |
| 57500 | 0.9293 | 49.0 | 120.5 | 12.0289 | 25.2405 | 99.047 | 12.401 | 27.375 | 34.5 |
| 60000 | 0.9697 | 48.75 | 120.5 | 12.0196 | 25.192 | 99.238 | 12.425 | 27.375 | 35.0 |
| 61875 | 1.0 | 49.0 | 121.0 | 12.0190 | 25.1853 | 99.264 | 12.428 | 27.375 | 35.0 |

# Resource Usage Comparison

- VRAM Use: 7.7823 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 145,731,804 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=25.0, loss_fn=cos, 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=25.0, loss_fn=cos, layer_mapper=layer-2))`
- train_embeddings: `True`
- lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7f040856c4f0>`
- 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