autoresearch
A small setup for agents to iterate on single-GPU language-model training experiments. It is a specialized research loop, not a general web-research assistant.
Utility: Research; kind: agent.
Stars: 98k+; Starred in September: 2k+.
Metrics updated: 31 Oct 2026. Profile reviewed: 31 Oct 2026.
Utilities
- Have a coding agent modify training code and record experimental results.
- Control an experiment loop through the repository's program instructions.
Who it is for
- ML researchers comfortable with training code and GPU experiments.
Requirements
- A single NVIDIA GPU, Python 3.10+, and uv are documented requirements.
- Provide a compatible coding agent and prepare the documented training data.
Limitations
- The current setup does not target CPU or Apple MPS execution.
- A short-run training metric is not evidence of general model improvement; inspect experiment validity.
Cost and privacy notes
- GPU runtime and coding-agent model usage can incur ongoing costs, especially unattended.
- Training data and logs need their own handling policy; the coding agent's provider may receive code and results.
Illustrative example (untested)
Illustrative, not tested: constrain an agent to one training-code change at a time and inspect its experiment log before retaining changes.
License
README declares MIT, but no standalone root license file was returned during review; verify the applicable grant before redistribution.