| |
| """Summarize a Harbor job (or trials directory) run on the MissionReport tasks. |
| |
| usage: python scripts/summarize_job.py <job or trials directory> [--criteria] |
| |
| Prints one line per trial: task, model, reward, positive points earned, penalty points triggered. |
| With --criteria, also lists each criterion the report missed and each penalty it triggered, with |
| the judge's one-line reason. Reads result.json and steps/answer/verifier/reward-details.json from |
| each trial directory. Standard library only. |
| """ |
| import json |
| import sys |
| from pathlib import Path |
|
|
|
|
| def criteria_of(trial: Path) -> list[dict]: |
| details = trial / "steps" / "answer" / "verifier" / "reward-details.json" |
| if not details.is_file(): |
| return [] |
| rubric = json.loads(details.read_text()).get("reward", {}).get("rubric") or {} |
| return rubric.get("criteria") or [] |
|
|
|
|
| def main() -> None: |
| root = Path(sys.argv[1]) |
| show = "--criteria" in sys.argv |
| rewards = [] |
| for result in sorted(root.rglob("result.json")): |
| trial = result.parent |
| data = json.loads(result.read_text()) |
| if "trial_name" not in data: |
| continue |
| task = data.get("task_name", trial.name).split("/")[-1] |
| model = ((data.get("agent_info") or {}).get("model_info") or {}).get("name") or data["agent_info"]["name"] |
| reward = ((data.get("verifier_result") or {}).get("rewards") or {}).get("reward") |
| error = (data.get("exception_info") or {}).get("exception_type") |
| criteria = criteria_of(trial) |
| earned = sum(c["signed_score"] for c in criteria if c["signed_score"] > 0 and c.get("raw")) |
| possible = sum(c["signed_score"] for c in criteria if c["signed_score"] > 0) |
| penalties = -sum(c["signed_score"] for c in criteria if c["signed_score"] < 0 and c.get("raw")) |
| shown = "error: " + error if error else ("no reward" if reward is None else f"{reward:.3f}") |
| print(f"{task:8} {model:32} {shown:>12} earned {earned:g}/{possible:g} penalties {penalties:g}") |
| if reward is not None: |
| rewards.append(reward) |
| if show: |
| for c in criteria: |
| missed = c["signed_score"] > 0 and not c.get("raw") |
| triggered = c["signed_score"] < 0 and c.get("raw") |
| if missed or triggered: |
| tag = "MISSED " if missed else "PENALTY" |
| print(f" {tag} {c['signed_score']:+g} {c['description']}") |
| if c.get("judge_reasoning"): |
| print(f" judge: {c['judge_reasoning']}") |
| if rewards: |
| print(f"\n{len(rewards)} graded trials, mean reward {sum(rewards) / len(rewards):.3f}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
