---
base_model: gpt2
datasets:
- wikimedia/wikipedia
library_name: Distily
license: mit
tags:
- bitnet
- 1.58b
- generated_from_trainer
model-index:
- name: distily_projector_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 | 22.0554 | 29.7695 | 83.978 | 10.514 | 4060086272.0 | 71468255805440.0 |
| 2500 | 0.0404 | 744.0 | 5792.0 | 2.6072 | 29.8269 | 83.817 | 10.494 | 450.0 | 4192.0 |
| 5000 | 0.0808 | 316.0 | 1408.0 | 1.9087 | 29.8408 | 83.778 | 10.489 | 237.0 | 300.0 |
| 7500 | 0.1212 | 221.0 | 776.0 | 1.6174 | 29.8221 | 83.83 | 10.496 | 180.0 | 175.0 |
| 10000 | 0.1616 | 165.0 | 616.0 | 1.4278 | 29.8115 | 83.86 | 10.499 | 144.0 | 150.0 |
| 12500 | 0.2020 | 124.5 | 484.0 | 1.1898 | 29.8065 | 83.874 | 10.501 | 107.0 | 146.0 |
| 15000 | 0.2424 | 108.0 | 452.0 | 1.0632 | 29.7962 | 83.903 | 10.505 | 92.5 | 109.0 |
| 17500 | 0.2828 | 91.5 | 352.0 | 0.9642 | 29.7907 | 83.919 | 10.507 | 75.5 | 119.0 |
| 20000 | 0.3232 | 79.0 | 302.0 | 0.8892 | 29.8022 | 83.886 | 10.503 | 68.0 | 688.0 |
| 22500 | 0.3636 | 72.5 | 236.0 | 0.7640 | 29.8032 | 83.883 | 10.502 | 59.0 | 90.5 |
| 25000 | 0.4040 | 68.0 | 202.0 | 0.7181 | 29.8737 | 83.686 | 10.477 | 51.75 | 85.5 |
| 27500 | 0.4444 | 63.75 | 223.0 | 0.6853 | 29.8142 | 83.853 | 10.498 | 48.0 | 99.0 |
| 30000 | 0.4848 | 63.5 | 214.0 | 0.6791 | 29.8122 | 83.858 | 10.499 | 50.75 | 76.5 |
| 32500 | 0.5253 | 64.0 | 194.0 | 0.6660 | 29.794 | 83.91 | 10.505 | 46.75 | 96.5 |
| 35000 | 0.5657 | 60.25 | 176.0 | 0.6075 | 29.8902 | 83.639 | 10.472 | 41.5 | 62.25 |
| 37500 | 0.6061 | 60.5 | 169.0 | 0.5942 | 29.8616 | 83.72 | 10.482 | 43.5 | 78.5 |
| 40000 | 0.6465 | 57.5 | 176.0 | 0.5808 | 29.8082 | 83.87 | 10.5 | 39.5 | 80.5 |
| 42500 | 0.6869 | 58.0 | 172.0 | 0.5602 | 29.7977 | 83.899 | 10.504 | 40.0 | 58.75 |
| 45000 | 0.7273 | 52.5 | 145.0 | 0.4723 | 29.7887 | 83.924 | 10.507 | 34.25 | 47.0 |
| 47500 | 0.7677 | 52.75 | 135.0 | 0.4507 | 29.7668 | 83.986 | 10.515 | 33.5 | 41.25 |
| 50000 | 0.8081 | 51.25 | 133.0 | 0.4370 | 29.7994 | 83.894 | 10.504 | 31.875 | 39.75 |
| 52500 | 0.8485 | 49.75 | 127.5 | 0.4272 | 29.7762 | 83.96 | 10.512 | 32.25 | 38.0 |
| 55000 | 0.8889 | 49.25 | 126.5 | 0.4130 | 29.814 | 83.853 | 10.498 | 31.25 | 35.5 |
| 57500 | 0.9293 | 48.5 | 125.0 | 0.4079 | 29.7893 | 83.923 | 10.507 | 30.625 | 34.25 |
| 60000 | 0.9697 | 48.75 | 123.5 | 0.4046 | 29.8263 | 83.819 | 10.494 | 30.625 | 34.75 |
| 61875 | 1.0 | 48.75 | 124.0 | 0.4043 | 29.85 | 83.752 | 10.486 | 30.625 | 34.5 |

# Resource Usage Comparison

- VRAM Use: 7.7843 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=10.0, loss_fn=raw_mse, 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=10.0, loss_fn=raw_mse, layer_mapper=layer-2))`
- train_embeddings: `True`
- lr_scheduler: `<torch.optim.lr_scheduler.LambdaLR object at 0x7f010c128160>`
- 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