---
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 | 25.7744 | 30.269 | 82.593 | 10.341 | 4060086272.0 | 71468255805440.0 |
| 2500 | 0.0404 | 952.0 | 8032.0 | 6.1202 | 30.2377 | 82.678 | 10.351 | 652.0 | 6432.0 |
| 5000 | 0.0808 | 378.0 | 1880.0 | 5.0293 | 30.3345 | 82.414 | 10.318 | 270.0 | 288.0 |
| 7500 | 0.1212 | 230.0 | 820.0 | 4.5127 | 30.2749 | 82.577 | 10.339 | 201.0 | 174.0 |
| 10000 | 0.1616 | 173.0 | 628.0 | 4.2288 | 30.3713 | 82.315 | 10.306 | 152.0 | 172.0 |
| 12500 | 0.2020 | 127.5 | 482.0 | 3.8554 | 30.4218 | 82.178 | 10.289 | 106.5 | 157.0 |
| 15000 | 0.2424 | 109.0 | 436.0 | 3.6684 | 30.2501 | 82.644 | 10.347 | 87.5 | 148.0 |
| 17500 | 0.2828 | 93.5 | 350.0 | 3.5226 | 30.2559 | 82.628 | 10.345 | 73.5 | 119.5 |
| 20000 | 0.3232 | 75.5 | 282.0 | 3.3351 | 30.5792 | 81.755 | 10.236 | 64.5 | 114.5 |
| 22500 | 0.3636 | 67.5 | 213.0 | 3.1522 | 30.4457 | 82.113 | 10.281 | 53.25 | 80.5 |
| 25000 | 0.4040 | 64.0 | 197.0 | 3.0806 | 30.4623 | 82.069 | 10.275 | 45.0 | 95.5 |
| 27500 | 0.4444 | 58.25 | 203.0 | 3.0308 | 30.2865 | 82.545 | 10.335 | 40.75 | 93.0 |
| 30000 | 0.4848 | 59.5 | 195.0 | 3.0207 | 30.2442 | 82.66 | 10.349 | 43.25 | 61.5 |
| 32500 | 0.5253 | 58.5 | 172.0 | 3.0015 | 30.3473 | 82.38 | 10.314 | 41.75 | 60.75 |
| 35000 | 0.5657 | 57.25 | 171.0 | 2.9425 | 30.2295 | 82.701 | 10.354 | 38.0 | 51.5 |
| 37500 | 0.6061 | 57.25 | 157.0 | 2.9166 | 31.0953 | 80.398 | 10.066 | 38.0 | 55.0 |
| 40000 | 0.6465 | 54.75 | 160.0 | 2.8987 | 30.9607 | 80.747 | 10.11 | 35.5 | 58.75 |
| 42500 | 0.6869 | 53.75 | 153.0 | 2.8783 | 30.2174 | 82.734 | 10.358 | 35.75 | 67.0 |
| 45000 | 0.7273 | 50.25 | 136.0 | 2.7764 | 30.1636 | 82.881 | 10.377 | 30.375 | 41.25 |
| 47500 | 0.7677 | 50.25 | 126.5 | 2.7509 | 30.252 | 82.639 | 10.346 | 29.375 | 38.0 |
| 50000 | 0.8081 | 49.0 | 127.5 | 2.7362 | 30.1988 | 82.785 | 10.365 | 28.75 | 40.0 |
| 52500 | 0.8485 | 48.5 | 121.0 | 2.7263 | 30.2772 | 82.57 | 10.338 | 29.0 | 35.5 |
| 55000 | 0.8889 | 48.0 | 119.0 | 2.7102 | 30.1534 | 82.909 | 10.38 | 28.0 | 34.0 |
| 57500 | 0.9293 | 47.5 | 118.5 | 2.7043 | 30.181 | 82.834 | 10.371 | 27.875 | 32.5 |
| 60000 | 0.9697 | 47.5 | 118.0 | 2.7013 | 30.3322 | 82.421 | 10.319 | 27.75 | 32.25 |
| 61875 | 1.0 | 47.5 | 118.0 | 2.7006 | 30.5306 | 81.885 | 10.252 | 27.625 | 32.25 |

# Resource Usage Comparison

- VRAM Use: 7.7830 GB

`# Distillation (Teacher -> Student) Architecture Difference:

- **Architecture**: `GPT2LMHeadModel` -> `GPT2LMHeadModel`
- **Total Parameters**: 124,439,808 -> 124,439,808
- **Data Type (dtype)**: 124439808 -> 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=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=5, loss_fn=cos, layer_mapper=layer-2))`
- train_embeddings: `True`
- lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7fb7a85a5a50>`
- 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 4.44.0
- Pytorch 2.3.0
- Datasets 2.21.0