Nanochat: training your own LLM chat bot from scratch

I’ll walk the audience through Andrej Karpathy’s latest reincarnation of minGPT/nanogpt, namely nanochat. This repo contains all major LLM stages including tokenization, pretraining, finetuning, evaluation, inference, and even a chat UI. Its minimalistic implementation allows you to train and evaluate an LLM from scratch, on a single GPU node or even a CPU. With access to better GPU machines (e.g. an 8X H100), you can get up to GPT2-level performance in under 2 hours!