Let's train an LLM into speaking like John Danaher, arguably one of the greatest BJJ coach.

Try asking for his guidance!

Simply launch the app with python app.py

John Danaher regularly write high quality posts on Instagram where he develops ideas around BJJ in his captions. We will download these captions and feed them into a pre-trained LLM for supervised fine-tuning. This work is done in notebooks/data_collector.ipynb notebook where we scrap instagram and store the relevant data into the data/danaher_data.csv file.

We use the mistral7B open-source model from MistralAI available on Hugging Face.

We utilize the QLoRA training to greatly reduce both the training computational cost and memory footprint. This modern technique is readily available in python through the peft librairy and will allow us to conduct the training on consumer hardware !

We build a simple gradio app to demo the model.

• Add possibility to choose an open-source LLM and lauch a fine-tuning job easily
• Instead of fine-tuning, explore RAG solutions
• Instead of fine-tuning, try to build an instruct-like dataset and compare different methods like RLHF and DPO

Simple instructions to get started running the notebooks and using the web-app.

1. Clone the repo

   git clone https://github.com/NicolasPllr1/bjj_coach.git

2. Install python dependancies

   pip install -r requirements.txt

Simply launch the app with python app.py