ProgramGym

Welcome to ProgramGym 👋

This is the home of ProgramGym, a project dedicated to training coding agents to rebuild real-world programs from scratch.

ProgramGym brings together Program-Env and Program-Agent. Program-Env transforms GitHub repositories into verifiable environments for program reprodcution from scratch. Using this novel pipeline, we obtain the ProgramGym-1648, which consist of 1,648 training environments spanning C, C++, Go, and Rust. Program-Agent follows a role-isolated analysis–implementation–review workflow, supported by compilation checks and failure recovery.

Feel free to explore our repositories and dataset, and join us in advancing the capabilities of coding agents! 🤩