File size: 6,703 Bytes
f8c91b8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 | #!/usr/bin/env python3
"""Class-wise metrics and confidence-resolution analysis (camera-ready additions).
Derives, offline from the shipped per-trial data:
1. Per-model confusion matrices (yes = positive class), per-class precision /
recall / F1, macro-F1, and balanced accuracy under the validated protocol,
plus pooled rows for the validated and naive protocols
-> writes ../tables/tcb_classwise.tex (Table: class-wise decomposition).
2. Confidence-resolution statistics: accuracy on trials with stated
confidence below 100 versus exactly 100 (pooled and per model), and the
pooled AUROC of stated confidence for predicting correctness.
3. Spearman rank correlation between panel accuracy and panel ECE.
Stdlib only; no endpoint, no network.
"""
import json
import pathlib
import statistics
HERE = pathlib.Path(__file__).resolve().parent
DATA = HERE.parent / "data"
TABLES = HERE.parent / "tables"
# Table 1 display order and short names.
ORDER = [
("glm-5.1-fp8", "glm"),
("deepseek-v4-pro", "deepseek"),
("gemma-4-31b", "gemma"),
("gptoss-120b", "gptoss"),
("leanstral-2603", "leanstral"),
("qwen3-omni-30b-a3b", "qwen-omni"),
("qwen3.6-35b-a3b-fp8", "qwen3.6-35b"),
("qwen3.6-27b-fp8", "qwen3.6-27b"),
("qwen3.5-9b", "qwen3.5-9b"),
]
def confusion(trials):
tp = fn = fp = tn = 0
for t in trials:
a = t.get("answer")
if a not in ("Yes", "No"):
continue
if t["truth"] == "Yes":
tp += a == "Yes"
fn += a == "No"
else:
fp += a == "Yes"
tn += a == "No"
return tp, fn, fp, tn
def classwise(tp, fn, fp, tn):
p_yes = tp / (tp + fp) if tp + fp else float("nan")
r_yes = tp / (tp + fn) if tp + fn else float("nan")
f_yes = 2 * p_yes * r_yes / (p_yes + r_yes) if p_yes + r_yes else 0.0
p_no = tn / (tn + fn) if tn + fn else float("nan")
r_no = tn / (tn + fp) if tn + fp else float("nan")
f_no = 2 * p_no * r_no / (p_no + r_no) if p_no + r_no else 0.0
return dict(p_yes=p_yes, r_yes=r_yes, f_yes=f_yes, p_no=p_no, r_no=r_no,
f_no=f_no, macro_f1=(f_yes + f_no) / 2, bal_acc=(r_yes + r_no) / 2)
def spearman(x, y):
def rank(v):
s = sorted(v)
return [(sum(1 for u in s if u < a) + (sum(1 for u in s if u == a) + 1) / 2)
for a in v]
rx, ry = rank(x), rank(y)
mx, my = statistics.mean(rx), statistics.mean(ry)
num = sum((a - mx) * (b - my) for a, b in zip(rx, ry))
den = (sum((a - mx) ** 2 for a in rx) * sum((b - my) ** 2 for b in ry)) ** 0.5
return num / den
def main():
trials = json.load(open(DATA / "tcb_v03_trials.json"))
naive = json.load(open(DATA / "tcb_v03_trials_naive.json"))
lines = [
"\\setlength{\\tabcolsep}{3.0pt}",
"\\begin{tabular}{lrrrrrrrrrrrr}",
"\\toprule",
" & \\multicolumn{4}{c}{Confusion} & \\multicolumn{3}{c}{Yes class}"
" & \\multicolumn{3}{c}{No class} & & \\\\",
"\\cmidrule(lr){2-5}\\cmidrule(lr){6-8}\\cmidrule(lr){9-11}",
"Model & TP & FN & FP & TN & Prec & Rec & F1 & Prec & Rec & F1"
" & Macro-F1 & Bal.\\ acc. \\\\",
"\\midrule",
]
pooled_tp = pooled_fn = pooled_fp = pooled_tn = 0
for model, short in ORDER:
tp, fn, fp, tn = confusion([t for t in trials if t["model"] == model])
pooled_tp += tp; pooled_fn += fn; pooled_fp += fp; pooled_tn += tn
m = classwise(tp, fn, fp, tn)
lines.append(
f"{short} & {tp} & {fn} & {fp} & {tn}"
f" & {m['p_yes']:.3f} & {m['r_yes']:.3f} & {m['f_yes']:.3f}"
f" & {m['p_no']:.3f} & {m['r_no']:.3f} & {m['f_no']:.3f}"
f" & {m['macro_f1']:.3f} & {m['bal_acc']:.3f} \\\\")
lines.append("\\midrule")
for label, (tp, fn, fp, tn) in (
("Pooled (validated)", (pooled_tp, pooled_fn, pooled_fp, pooled_tn)),
("Pooled (naive)", confusion(naive))):
m = classwise(tp, fn, fp, tn)
lines.append(
f"{label} & {tp} & {fn} & {fp} & {tn}"
f" & {m['p_yes']:.3f} & {m['r_yes']:.3f} & {m['f_yes']:.3f}"
f" & {m['p_no']:.3f} & {m['r_no']:.3f} & {m['f_no']:.3f}"
f" & {m['macro_f1']:.3f} & {m['bal_acc']:.3f} \\\\")
lines += ["\\bottomrule", "\\end{tabular}"]
out = TABLES / "tcb_classwise.tex"
out.write_text("\n".join(lines) + "\n")
print(f"wrote {out}")
# Confidence resolution: does below-ceiling stated confidence predict error?
stated = [t for t in trials
if t.get("conf_stated") and t.get("answer") in ("Yes", "No")]
at100 = [t for t in stated if t["conf"] == 100]
below = [t for t in stated if t["conf"] < 100]
acc = lambda ts: sum(t["correct"] for t in ts) / len(ts)
print(f"stated-confidence trials: {len(stated)}"
f" (of {sum(1 for t in trials if t.get('answer') in ('Yes','No'))} parseable)")
print(f"conf = 100: n={len(at100)} acc={acc(at100):.3f}")
print(f"conf < 100: n={len(below)} acc={acc(below):.3f}")
for model, short in ORDER:
lo = [t for t in below if t["model"] == model]
hi = [t for t in at100 if t["model"] == model]
if len(lo) >= 10:
print(f" {short}: conf<100 n={len(lo)} acc={acc(lo):.3f}"
f" vs conf=100 acc={acc(hi):.3f}")
correct = [t["conf"] for t in stated if t["correct"]]
wrong = [t["conf"] for t in stated if not t["correct"]]
wins = sum(1 for a in correct for b in wrong if a > b)
ties = sum(1 for a in correct for b in wrong if a == b)
auroc = (wins + 0.5 * ties) / (len(correct) * len(wrong))
print(f"pooled AUROC(stated conf -> correct) = {auroc:.3f}"
f" (n_correct={len(correct)}, n_incorrect={len(wrong)})")
# Accuracy vs ECE rank correlation across the panel (Table 1 values are
# re-derivable via analyze_panel.py; recomputed here from the trials).
accs, eces = [], []
for model, short in ORDER:
ts = [t for t in trials if t["model"] == model
and t.get("answer") in ("Yes", "No")]
accs.append(acc(ts))
confs = [(t["conf"] if t.get("conf_stated") else 50) / 100 for t in ts]
bins = [[] for _ in range(5)]
for c, t in zip(confs, ts):
bins[min(4, int(c * 5))].append((c, t["correct"]))
ece = sum(len(b) / len(ts) *
abs(statistics.mean(c for c, _ in b) -
statistics.mean(k for _, k in b))
for b in bins if b)
eces.append(ece)
print(f"Spearman(accuracy, ECE) across the panel: {spearman(accs, eces):.3f}")
if __name__ == "__main__":
main()
|