E2E LLM finetuning on a single 24GB GPU

I have experience working with Large Language Models (LLMs) during my tenure at my previous company. However, one aspect I didn't get the opportunity to explore there was fine-tuning my own LLM, so I decided to experiment.

While there are many pieces of documentation available on this topic scattered across the Internet, putting them all together is not obvious. Therefore, in this repository, I present the full training code, from scrapping, to generating the dataset to the finetuning itself.

The whole thing takes a few days to complete, and it does work on my AMD GPU!