scripts/TRAIN.md

# Train

## Environment

```bash
cd scripts
python -m venv venv
source venv/bin/activate
pip install -U -r requirements.in
```

## Tokenizer

```bash
python -B train_tokenizer.py
```

## Dataset

```bash
python -B prepare_pretrain_dataset.py
```

## Model

### Pretrain

```bash
litgpt pretrain --config ./pretrain-model.yaml
```

```bash
litgpt convert_from_litgpt out/pretrain/final/ out/converted_model
cp config.json out/pretrain/final/
cp config.json out/converted_model/
```

```python
import torch
from safetensors.torch import save_file

state_dict = torch.load('out/converted_model/model.pth', map_location='cpu')
save_file(state_dict, 'out/converted_model/model.safetensors')
```

## Evaluate

```bash
litgpt evaluate --tasks 'hellaswag,gsm8k,truthfulqa_mc2,mmlu,winogrande,arc_challenge' --out_dir 'evaluate-quick/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/

litgpt evaluate --tasks 'leaderboard' --out_dir 'evaluate-leaderboard/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/

litgpt evaluate --tasks 'bbh_zeroshot,bbh_fewshot,bbh_cot_fewshot,bbh_cot_zeroshot' --out_dir 'evaluate-bigbenchhard/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/

litgpt evaluate --tasks 'mmlu,mmlu_pro' --out_dir 'evaluate-mmlu/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/

litgpt evaluate --tasks 'arc_challenge,boolq,gpqa,hellaswag,openbookqa,piqa,siqa,truthfulqa_mc2,winogrande' --out_dir 'evaluate-reasoning/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/

litgpt evaluate --tasks 'mmlu_multilingual,mgsm' --out_dir 'evaluate-multilinguals/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/

litgpt evaluate --tasks 'gsm8k,mathqa' --out_dir 'evaluate-math/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/

litgpt evaluate --tasks 'qasper' --out_dir 'evaluate-long/' --batch_size 4 --dtype 'bfloat16' out/pretrain/final/
```