autoresearch

About

autoresearch is a pioneering AI agent research tool by karpathy, designed for autonomous exploration and optimization of Large Language Model (LLM) training. It empowers an AI agent to modify a single file (`train.py`), which includes model architecture, hyperparameters, and training logic. The agent conducts fixed 5-minute training runs, evaluating performance using `val_bpb` as the core metric for iterative self-improvement. Guided by human-programmed instructions in `program.md`, this system aims to automatically discover and refine optimal LLM configurations on a single-GPU setup, significantly accelerating research and enabling unattended, overnight experimental breakthroughs.

Supported Platforms