Datasets:
File size: 1,788 Bytes
c3cd245 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 | """Score one LegalReasonBench matter from a finished Harbor trial.
The matter reward is points-weighted across its assignments: the sum of signed
points earned (penalties subtract) over the sum of available positive points,
clamped to [0, 1]. Harbor's own multi_step_reward_strategy = "mean" averages the
per-assignment rewards instead and is diagnostic only.
Usage: python score_matter.py <harbor trial dir>
"""
import json
import sys
from pathlib import Path
def score(trial_dir: Path) -> dict:
earned = positive = 0.0
steps = []
# Each assignment's verifier writes reward-details.json; the raw RewardKit copy
# under raw/ has no aggregation block and is skipped.
for details in sorted(trial_dir.rglob("reward-details.json")):
correctness = json.loads(details.read_text()).get("correctness")
aggregation = correctness.get("aggregation") if isinstance(correctness, dict) else None
if not isinstance(aggregation, dict) or "earned" not in aggregation or "positive_total" not in aggregation:
continue
earned += float(aggregation["earned"])
positive += float(aggregation["positive_total"])
steps.append({
"file": str(details.relative_to(trial_dir)),
"assignment_reward": correctness.get("score"),
"earned": aggregation["earned"],
"positive_total": aggregation["positive_total"],
})
if not steps:
raise SystemExit(f"no assignment reward-details.json under {trial_dir}")
overall = max(0.0, min(1.0, earned / positive)) if positive > 0 else 0.0
return {"matter_reward": overall, "earned": earned, "positive_total": positive, "assignments": steps}
if __name__ == "__main__":
print(json.dumps(score(Path(sys.argv[1])), indent=2))
|