---
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 | 30.2402 | 82.672 | 10.35 | 4060086272.0 | 71468255805440.0 |
| 2500 | 0.0404 | 940.0 | 7712.0 | 6.1058 | 30.4416 | 82.125 | 10.282 | 640.0 | 6272.0 |
| 5000 | 0.0808 | 378.0 | 1880.0 | 5.0293 | 30.4354 | 82.141 | 10.284 | 270.0 | 288.0 |
| 7500 | 0.1212 | 230.0 | 820.0 | 4.5127 | 30.4162 | 82.193 | 10.291 | 201.0 | 174.0 |
| 10000 | 0.1616 | 173.0 | 632.0 | 4.2293 | 30.2401 | 82.672 | 10.351 | 151.0 | 172.0 |
| 12500 | 0.2020 | 127.5 | 482.0 | 3.8556 | 30.2143 | 82.742 | 10.359 | 106.5 | 156.0 |
| 15000 | 0.2424 | 109.0 | 436.0 | 3.6684 | 30.2343 | 82.688 | 10.352 | 87.5 | 144.0 |
| 17500 | 0.2828 | 93.5 | 348.0 | 3.5229 | 30.333 | 82.419 | 10.319 | 73.5 | 122.5 |
| 20000 | 0.3232 | 73.5 | 276.0 | 3.3349 | 30.3063 | 82.491 | 10.328 | 63.25 | 99.5 |
| 22500 | 0.3636 | 67.0 | 219.0 | 3.1509 | 30.3619 | 82.34 | 10.309 | 52.25 | 79.0 |
| 25000 | 0.4040 | 64.5 | 189.0 | 3.0823 | 30.4079 | 82.215 | 10.293 | 45.75 | 97.0 |
| 27500 | 0.4444 | 59.0 | 194.0 | 3.0271 | 30.4181 | 82.188 | 10.29 | 41.25 | 85.5 |
| 30000 | 0.4848 | 59.25 | 194.0 | 3.0192 | 30.2505 | 82.643 | 10.347 | 42.75 | 57.75 |
| 32500 | 0.5253 | 58.5 | 175.0 | 3.0025 | 30.2733 | 82.581 | 10.339 | 40.0 | 62.75 |
| 35000 | 0.5657 | 57.0 | 170.0 | 2.9448 | 30.2658 | 82.601 | 10.342 | 37.0 | 54.25 |
| 37500 | 0.6061 | 57.25 | 155.0 | 2.9182 | 30.2187 | 82.73 | 10.358 | 38.75 | 73.5 |
| 40000 | 0.6465 | 54.75 | 164.0 | 2.8978 | 30.2683 | 82.595 | 10.341 | 35.25 | 70.0 |
| 42500 | 0.6869 | 54.25 | 156.0 | 2.8775 | 30.4126 | 82.203 | 10.292 | 34.75 | 61.75 |
| 45000 | 0.7273 | 50.25 | 137.0 | 2.7761 | 30.3396 | 82.401 | 10.317 | 30.5 | 60.75 |
| 47500 | 0.7677 | 50.25 | 126.5 | 2.7499 | 30.3808 | 82.289 | 10.303 | 29.5 | 37.25 |
| 50000 | 0.8081 | 49.25 | 126.5 | 2.7359 | 30.3056 | 82.493 | 10.328 | 28.625 | 37.75 |
| 52500 | 0.8485 | 48.5 | 122.0 | 2.7258 | 30.3024 | 82.502 | 10.329 | 29.125 | 36.25 |
| 55000 | 0.8889 | 48.0 | 119.0 | 2.7099 | 30.201 | 82.779 | 10.364 | 28.125 | 34.0 |
| 57500 | 0.9293 | 47.5 | 119.0 | 2.7046 | 30.1798 | 82.837 | 10.371 | 27.875 | 33.5 |
| 60000 | 0.9697 | 47.75 | 118.5 | 2.7011 | 30.355 | 82.359 | 10.311 | 27.75 | 33.0 |
| 61875 | 1.0 | 47.75 | 119.0 | 2.7006 | 30.4772 | 82.028 | 10.27 | 27.875 | 33.0 |

# Resource Usage Comparison

- VRAM Use: 7.7831 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 0x7fbd2875e9e0>`
- 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