karpathy/nanochat ? reverse-engineered prompt

I want to build a simple setup for training my own small language models, based on Andrej Karpathy's nanochat project. The main idea is to reproduce the "GPT-2 speedrun" that's described, which should let me train a custom chatbot for a relatively low cost on a single powerful GPU node.

Please create a complete Python project that handles the entire pipeline: data tokenization, model pretraining, and then running inference. After the model is finished training, which I know might take a few hours, I want to be able to launch a simple web UI. This web interface should look like a basic version of ChatGPT, where I can type messages and get responses from the model I just trained.

Set up the project with all the dependencies using uv, as the repo suggests. The end goal is to have a system where I can kick off the training process, and once it's done, easily start the chat server to talk to my new AI.