PostTrainBench

Why we included this project

PostTrainBench offers a concrete way to test whether an autonomous coding agent can do real model development work. It hands a CLI agent like Claude Code, Codex, Gemini CLI, or OpenCode a base LLM, an evaluation script, and ten hours on a single H100, then scores how much the agent improves the model on a target benchmark. The seven tasks span math, code, tool use, writing, and medical reasoning, so the results say something about general research ability rather than one narrow skill. Teams comparing agentic tooling for research automation will appreciate that every agent runs under identical conditions, and the included baselines and evaluation scripts make runs easy to reproduce. It is a research harness, not a turnkey product: expect to work with containers and a job scheduler, though the design is clean enough to adapt to your own hardware.