benchmax

Design a castform RL environment — a BaseEnv subclass with tools and reward functions. Use when creating or editing main.py / the env for a training run.

Create the train/eval datasets for a castform run — generic prompt/ground_truth jsonl, or RAG QA pairs from a corpus (`corpus ingest` + `data qa-gen`). Use when building or uploading training data.

Launch a castform GPU training run with `castform launch` (validate → upload → launch) and set launcher args correctly. Use only after verify-environment is green — this spends GPU.

Verify a castform env with `castform validate` and interpret the rewards/errors before spending GPU. Use after editing the env or data, and before launch-run.

Monitor a castform training run — status, reward scalars, and logs via the castform runs commands. Use after launching to watch a run or debug a stalled/failed one.