File size: 3,293 Bytes
329b63d | 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 | #!/usr/bin/env python3
"""ChainCheck evaluator (standalone; needs only numpy).
Score every item of a ChainCheck file with your evidence-sufficiency system (higher = "the passages are
sufficient"), then:
python chaincheck_eval.py --data twowiki_replication.jsonl.gz --scores my_scores.jsonl [--exclude-cb27b]
my_scores.jsonl: one JSON object per line, {"item_id": ..., "score": float}.
Reports, separately for real and synthetic replacement entities:
Nominal = AUC(A > D) what an ordinary benchmark reports
CE = AUC(B > D) - 0.5 chain effect: response to breaking the chain, edit held fixed
EE = AUC(A > B) - 0.5 edit effect: response to an edit that leaves the chain intact
Sigma = CE - |EE| chain selectivity; chain-selective if the 95% lower bound > 0
Pair-bootstrap 95% intervals (2,000 resamples, seed 0). These are matched contrasts, not an additive
decomposition: Nominal is not CE + EE.
--exclude-cb27b drops pairs whose question Qwen3.8-27B answered closed-book (the paper's closed-book-hard filter).
"""
import argparse, gzip, json
from collections import defaultdict
import numpy as np
def load(path):
op = gzip.open if path.endswith(".gz") else open
with op(path, "rt", encoding="utf-8") as f:
return [json.loads(x) for x in f if x.strip()]
def wins(s, hi, lo):
return np.array([(s[h] > s[l]) + 0.5 * (s[h] == s[l]) for h, l in zip(hi, lo)], dtype=float)
def metrics(s, quads, n_boot=2000, seed=0):
A = [q["A"] for q in quads]; B = [q["B"] for q in quads]; D = [q["D"] for q in quads]
W = {"nominal": wins(s, A, D), "chain": wins(s, B, D), "edit": wins(s, A, B)}
idx = np.random.default_rng(seed).integers(0, len(quads), size=(n_boot, len(quads)))
ce_b = W["chain"][idx].mean(1) - 0.5; ee_b = W["edit"][idx].mean(1) - 0.5; sig_b = ce_b - np.abs(ee_b)
ci = lambda x: [round(float(np.percentile(x, 2.5)), 4), round(float(np.percentile(x, 97.5)), 4)]
ce, ee = W["chain"].mean() - 0.5, W["edit"].mean() - 0.5
return {"n_pairs": len(quads), "nominal": round(float(W["nominal"].mean()), 4), "CE": round(float(ce), 4),
"CE_ci": ci(ce_b), "EE": round(float(ee), 4), "EE_ci": ci(ee_b), "sigma": round(float(ce - abs(ee)), 4),
"sigma_ci": ci(sig_b), "chain_selective": bool(np.percentile(sig_b, 2.5) > 0)}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--data", required=True); ap.add_argument("--scores", required=True)
ap.add_argument("--exclude-cb27b", action="store_true")
a = ap.parse_args()
rows = load(a.data)
sc = {x["item_id"]: float(x["score"]) for x in load(a.scores)}
missing = [r["item_id"] for r in rows if r["item_id"] not in sc]
if missing:
raise SystemExit(f"{len(missing)} items have no score, e.g. {missing[:3]}")
quads = defaultdict(dict)
for r in rows:
if a.exclude_cb27b and r["cb_27b_correct"]:
continue
quads[(r["pair_id"], r["variant"])][r["cell"]] = r["item_id"]
out = {}
for variant in ("real", "fict"):
qs = [q for (p, v), q in sorted(quads.items()) if v == variant and len(q) == 3]
if qs:
out[variant] = metrics(sc, qs)
print(json.dumps(out, indent=1))
if __name__ == "__main__":
main()
|