traindoctor-env

Executable RL environment for diagnosing and repairing distributed PyTorch training failures, with sandboxed execution and correctness-verified recovery rewards.

TrainDoctor is an executable RL environment in which an agent repairs a failed multi-process PyTorch training job. Each candidate repair is re-executed in a network-less Modal (gVisor) sandbox, and a verifier outside the sandbox checks it against records written by the launcher rather than by the job. For example, setting world_size: 1 removes a collective hang and the job completes, but in one red-team case 72 scheduled samples were never trained on, 8 were trained on twice, and the model state differed from the reference; the repair is rejected.