LLM-Finetune

A demo to finetune LLM using DeepSpeed x HuggingFace Trainer.

Usage

1. Prepare the environment. Install dependencies listed in the pyproject.toml file. You may need to install OpenMPI beforehand.
2. Log into Weights & Biases.
3. Prepare the config file. You can skip this if you use existing config files. If you prepare your own config file, first create a DeepSpeed config from config/deepspeed. Then specify the path to the DeepSpeed config in a main config.
4. Run training. When training with DeepSpeed, you can use deepspeed src/llm_finetune/train.py --config_path path/to/the/config.yaml. When training without DeepSpeed, you can use python src/llm_finetune/train.py --config_path path/to/the/config.yaml.
5. Have some fun with the trained model.

Demo: Finetuning GPT-2 with COVID-19 Papers

In the demo, I finetuned GPT-2 with Cord19 dataset, which is a corpus of research papers related to COVID-19. Due to the restriction in time and computational resource, I only trained it on very small samples. But the results shows that the model have aquired some knowledge on COVID-19. Check out the notebook/gpt2_results.ipynb for some comparison with original and finetuned GPT-2.