autoresearch
autoresearch - AI Agents for LLM Training Experiments
Quick facts
- Best for
- autoresearch - AI Agents for LLM Training Experiments
- Pricing
- Free
- Editor rating
- 4.5 / 5
- Community saves
- 0
About autoresearch
autoresearch is an MIT-licensed experimental repository for running AI-agent-driven research loops on a simplified single-GPU nanochat training setup. It uses a human-edited program.md brief, bounded training runs, and validation bits per byte to evaluate changes.
Pros
- Agent-oriented loop where code changes are proposed, run, measured, and accepted or rejected
- Fixed five-minute training budget for comparable nanochat experiments
- program.md file for human-written operating instructions to the coding agent
- Validation bits-per-byte metric for evaluating training changes
- Small Python codebase designed to be inspected and modified directly
Cons
Pricing
Open source
$0
- • MIT-licensed repository
- • No platform fee
- • Requires your own NVIDIA GPU or compatible compute environment