Environment setup

1. Clone or download this repo. cd yourself to it's root directory.
2. Create and activate python conda enviromnent: conda create --name data-distil python=3.8
3. Activate conda environment: conda activate data-distil
4. Install dependencies, using pip install -r requirements.txt
5. Set huggingface and openai credentials in .env

Lion Data Generation

1. Collect SLM's and LLM's predictions by running: python generation/seed_inference.py
2. Run: python generation/lion/generate.py (for slm_preds_path and llm_preds_path use save paths from Step 1.)
   • this will generate the following datasets (<SAVE_DIR> is specified in generate.py script, <DATE_TIME> is an automatically generated data id):
     • <SAVE_DIR>/<DATE_TIME>/lion_all
     • <SAVE_DIR>/<DATE_TIME>/lion_hard
     • <SAVE_DIR>/<DATE_TIME>/lion_easy

Finetuning and Inference

Supported SLMs:

• Mistral7B
• Llama3

1. Run: pyhton finetuning/pipeline.py
   • for finetuning on the Lion data, set ds_train param to the path of one of the locally saved Lion datasets, e.g. ds_train = "data/gsm8k/lion/2024-07-06_18-20-59/lion_hard" (here <SAVE_DIR> = "data/gsm8k/lion" and <DATE_TIME> = "2024-07-06_18-20-59")