Note Harbor env — 5,000 verified tasks for training. Agent explores a real dataset with a shell tool and answers; deterministic grading, no LLM judge.

Note Held-out benchmark — 250 verified tasks, difficulty-balanced and skewed harder to separate models. Report pass@1 / pass@k.

Note Validation split — 144 verified tasks for quick during-training checkpoints.

Note The same tasks as a plain, load_dataset-ready dataset (train/test/eval) — question, gold answer, grading rules and data location per row.

Note 4,677 verified (reward=1) tool-use trajectories for SFT. TRL-ready: conversational messages + tools.