File size: 4,229 Bytes
f10cac2 | 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 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 | #!/usr/bin/env python3
"""Keep only eval questions that BOTH king and challenger failed.
Reads scoring-results.jsonl. For each sample, a question is kept when every
parsed judge answered 0 for previous_king and 0 for challenger.
Usage:
python filter_shared_fails.py evals/<eval_run_id>
python filter_shared_fails.py evals/<eval_run_id>/scoring-results.jsonl
python filter_shared_fails.py evals/<eval_run_id> --out shared-fails.jsonl
"""
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
def load_records(path: Path) -> list[dict]:
if path.is_dir():
path = path / "scoring-results.jsonl"
if not path.is_file():
raise FileNotFoundError(f"no scoring-results.jsonl at {path}")
return [json.loads(line) for line in path.read_text().splitlines() if line.strip()]
def judge_by_side(record: dict) -> dict[str, dict]:
out: dict[str, dict] = {}
for result in record.get("judge_results") or []:
if not result.get("parse_ok"):
continue
side = result.get("side")
if side in ("previous_king", "challenger"):
out[side] = result
return out
def shared_fail_rows(record: dict) -> list[dict]:
judges = judge_by_side(record)
king = judges.get("previous_king")
chal = judges.get("challenger")
if not king or not chal:
return []
rows = []
for question in record.get("questions") or []:
qid = question.get("id")
ka = (king.get("answers") or {}).get(qid)
ca = (chal.get("answers") or {}).get(qid)
if ka != "0" or ca != "0":
continue
rows.append(
{
"sample_id": record.get("sample_id"),
"qid": qid,
"tag": question.get("tag"),
"weight": question.get("weight"),
"requires": question.get("requires"),
"step": question.get("step"),
"text": question.get("text"),
"example_bad": question.get("example_bad"),
"evidence": question.get("evidence"),
"king_answer": ka,
"chal_answer": ca,
"king_explanation": (king.get("explanations") or {}).get(qid),
"chal_explanation": (chal.get("explanations") or {}).get(qid),
"king_score": record.get("king_score"),
"challenger_score": record.get("challenger_score"),
}
)
return rows
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(
description="Filter scoring-results.jsonl to questions both models answered 0."
)
parser.add_argument("path", type=Path, help="eval directory or scoring-results.jsonl")
parser.add_argument(
"--out",
type=Path,
help="output jsonl (default: <eval-dir>/shared-fails.jsonl)",
)
args = parser.parse_args(argv)
source = args.path if args.path.is_file() else args.path / "scoring-results.jsonl"
records = load_records(args.path)
out_path = args.out or source.with_name("shared-fails.jsonl")
kept = 0
samples_with_fails = 0
samples_scored = 0
questions_total = 0
out_path.parent.mkdir(parents=True, exist_ok=True)
with out_path.open("w", encoding="utf-8") as handle:
for record in records:
questions_total += len(record.get("questions") or [])
if record.get("scored"):
samples_scored += 1
rows = shared_fail_rows(record)
if not rows:
continue
samples_with_fails += 1
for row in rows:
handle.write(json.dumps(row, ensure_ascii=False) + "\n")
kept += 1
print(f"scored samples {samples_scored}/{len(records)}")
print(f"questions {questions_total}")
print(f"shared fails {kept}")
print(f"samples with ≥1 {samples_with_fails}")
print(f"wrote {out_path}")
return 0
if __name__ == "__main__":
try:
raise SystemExit(main())
except Exception as exc: # noqa: BLE001
print(f"error: {exc}", file=sys.stderr)
raise SystemExit(1)
|