karpathy/nanochat

nanochat is a small experimental codebase for training and running language models, covering tokenization, pretraining, fine-tuning, evaluation, and inference. Researchers and developers use it to train GPT-style models on a single GPU node and interact with them through a command line.