commitment-test-llm-hallucination / code /analyze_study2_original.py
rejaulkorimsadi's picture
Release 2026-11-17
ca7f04f verified
Raw History Blame Contribute Delete
5.44 kB
#!/usr/bin/env python3
"""
run_analysis.py — computes Cohen's kappa and the H1/H2/H3 hypothesis tests
from data/annotator_A.csv and data/annotator_B.csv.
Reproduces kappa_results.json and h1_h2_h3_results.json exactly.
USAGE:
pip install scikit-learn scipy statsmodels
python run_analysis.py
"""
import csv
import json
from sklearn.metrics import cohen_kappa_score
from statsmodels.stats.contingency_tables import mcnemar
from statsmodels.stats.proportion import proportion_confint, proportions_ztest
from statsmodels.stats.multitest import multipletests
KAPPA_FLOOR = 0.80
def load(fname):
return {r["row_uid"]: r for r in csv.DictReader(open(fname, encoding="utf-8"))}
def as_num(s):
s = s.strip()
return float(s) if s else None
def main():
ann_a = load("data/annotator_A.csv")
ann_b = load("data/annotator_B.csv")
assert set(ann_a.keys()) == set(ann_b.keys()), "row_uid sets don't match"
print(f"Merged on {len(ann_a)} matching row_uids.\n")
# ---- Kappa ----
def kappa_for(rows, col):
s_vals = [as_num(ann_a[u][col]) for u in rows]
t_vals = [as_num(ann_b[u][col]) for u in rows]
agree = sum(1 for a, b in zip(s_vals, t_vals) if a == b) / len(rows)
try:
k = cohen_kappa_score(s_vals, t_vals)
except Exception:
k = float("nan")
return k, agree, len(rows)
print("=== COHEN'S KAPPA ===")
rows_iso = [u for u in ann_a if ann_a[u]["condition"] == "c_isolated"]
rows_sc = [u for u in ann_a if ann_a[u]["condition"] == "a_baseline"]
rows_ab = [u for u in ann_a if ann_a[u]["condition"] in ("b_substituted", "d_control")]
rows_nf = [u for u in ann_a if ann_a[u]["condition"] in ("a_baseline", "b_substituted", "d_control")]
for label, rows, col in [
("isolated_query_correct", rows_iso, "isolated_query_correct"),
("spontaneous_correction", rows_sc, "spontaneous_correction"),
("abstain", rows_ab, "abstain"),
]:
k, agree, n = kappa_for(rows, col)
print(f" {label}: n={n} agreement={agree:.1%} kappa={k}")
# n_downstream_failing dichotomized separately (needs >0 vs =0 transform)
s_vals = [1 if as_num(ann_a[u]["n_downstream_failing"]) > 0 else 0 for u in rows_nf]
t_vals = [1 if as_num(ann_b[u]["n_downstream_failing"]) > 0 else 0 for u in rows_nf]
agree = sum(1 for a, b in zip(s_vals, t_vals) if a == b) / len(rows_nf)
k = cohen_kappa_score(s_vals, t_vals)
print(f" n_downstream_failing (dichotomized): n={len(rows_nf)} agreement={agree:.1%} kappa={k:.3f}\n")
# ---- H1/H2/H3, using annotator A's scores as primary ----
primary = ann_a
by_item = {}
for uid, r in primary.items():
by_item.setdefault(r["item_id"], {})[r["condition"]] = r
b00 = b01 = b10 = b11 = 0
for it_id, conds in by_item.items():
if "a_baseline" not in conds or "b_substituted" not in conds:
continue
a = as_num(conds["a_baseline"]["n_downstream_failing"])
b = as_num(conds["b_substituted"]["n_downstream_failing"])
if a is None or b is None:
continue
a_err, b_err = int(a > 0), int(b > 0)
if a_err and b_err: b11 += 1
elif a_err and not b_err: b10 += 1
elif not a_err and b_err: b01 += 1
else: b00 += 1
pairs = b00 + b01 + b10 + b11
res_h1 = mcnemar([[b11, b10], [b01, b00]], exact=True)
rate_before = (b11 + b10) / pairs
rate_after = (b11 + b01) / pairs
print("=== H1 — COMMITMENT ===")
print(f" n={pairs} before={rate_before:.1%} after={rate_after:.1%} "
f"reduction={((rate_before-rate_after)*100):.1f}pts p={res_h1.pvalue:.6f}\n")
succeeded = [it for it, c in by_item.items()
if all(k in c for k in ("a_baseline", "b_substituted", "c_isolated"))
and as_num(c["a_baseline"]["n_downstream_failing"]) > 0
and as_num(c["b_substituted"]["n_downstream_failing"]) == 0]
iso_ok = sum(1 for it in succeeded if as_num(by_item[it]["c_isolated"]["isolated_query_correct"]) == 1)
n_succ = len(succeeded)
ci = proportion_confint(iso_ok, n_succ, method="wilson") if n_succ else (float("nan"),) * 2
print("=== H2 — KNOWLEDGE CHECK ===")
print(f" n={n_succ} correct_in_isolation={iso_ok} rate={iso_ok/n_succ:.1%} "
f"wilson_ci=[{ci[0]:.1%}, {ci[1]:.1%}]\n")
elim_B = elim_D = n_B = n_D = 0
for it, c in by_item.items():
if "b_substituted" in c:
v = as_num(c["b_substituted"]["n_downstream_failing"])
if v is not None:
n_B += 1
elim_B += v == 0
if "d_control" in c:
v = as_num(c["d_control"]["n_downstream_failing"])
if v is not None:
n_D += 1
elim_D += v == 0
z, p_h3 = proportions_ztest([elim_B, elim_D], [n_B, n_D])
print("=== H3 — CONTROL ===")
print(f" correct-sub: {elim_B}/{n_B}={elim_B/n_B:.1%} control-sub: {elim_D}/{n_D}={elim_D/n_D:.1%} "
f"z={z:.3f} p={p_h3:.6f}\n")
reject, q, _, _ = multipletests([res_h1.pvalue, p_h3], alpha=0.05, method="fdr_bh")
print("=== BENJAMINI-HOCHBERG (H1 + H3) ===")
print(f" H1: q={q[0]:.6f} {'SIGNIFICANT' if reject[0] else 'not significant'}")
print(f" H3: q={q[1]:.6f} {'SIGNIFICANT' if reject[1] else 'not significant'}")
if __name__ == "__main__":
main()