Datasets:
File size: 2,085 Bytes
fa0c0b8 | 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 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | """MA-Bench scorer. Standalone: python score.py rubric_items.jsonl verdicts.jsonl
verdicts.jsonl: one line per graded brief, {"case_id": "C1", "brief_id": "...", "failed_items": ["S1-01-n", ...]}
where failed_items are the rubric item_ids the judge marked as NOT satisfying pass_condition (the judge returns
failures only; every unlisted item counts as passed). For D and CD items the pass condition asserts the mistake is
present, so an unlisted (passed) negative item incurs its negative weight.
Prints per-brief unweighted pass rate and severity-weighted score, then per-case means.
"""
import json
import sys
import collections
import statistics as st
def load_items(path):
by = collections.defaultdict(dict)
for line in open(path):
it = json.loads(line)
by[it["case_id"]][it["item_id"]] = it
return by
def score(items, failed):
pos = [i for i in items.values() if i["weight"] > 0]
neg = [i for i in items.values() if i["weight"] < 0]
earned = sum(i["weight"] for i in pos if i["item_id"] not in failed)
incurred = sum(
i["weight"] for i in neg if i["item_id"] not in failed
) # mistake present
correct = sum(1 for i in pos if i["item_id"] not in failed) + sum(
1 for i in neg if i["item_id"] in failed
)
return {
"pass_rate": correct / len(items),
"weighted": (earned + incurred) / sum(i["weight"] for i in pos),
"detrimental_present": sum(1 for i in neg if i["item_id"] not in failed),
}
def main():
items = load_items(sys.argv[1])
per_case = collections.defaultdict(list)
for line in open(sys.argv[2]):
v = json.loads(line)
s = score(items[v["case_id"]], set(v["failed_items"]))
per_case[v["case_id"]].append(s)
print(
f"{v['case_id']}\t{v.get('brief_id', '')}\tpass={s['pass_rate']:.3f}\tweighted={s['weighted']:.3f}\tdetrimental={s['detrimental_present']}"
)
for c, ss in sorted(per_case.items()):
print(
f"# {c}\tn={len(ss)}\tpass={st.mean(x['pass_rate'] for x in ss):.3f}\tweighted={st.mean(x['weighted'] for x in ss):.3f}"
)
if __name__ == "__main__":
main()
|