autoresearch

Why we included this project

This repo hands an LLM agent a real, simplified single-GPU training loop and lets it run experiments overnight. The agent edits the model code, trains for five minutes, checks whether validation bits-per-byte improved, and keeps or discards the change before trying again. The design is refreshingly legible: one file the agent is allowed to touch, one file the human edits to steer it, and one metric to judge progress. It is a working experiment and a template for building your own agent loop rather than a production service, which is exactly the point. The README also covers scaling the setup down to smaller GPUs or Macs, useful if you do not have an H100 on hand.