| """Offline re-analyzer for the Q4 transfer study (Section 5.5). |
| |
| Recomputes, from the shipped per-trial files with no endpoint and no HuggingFace download, the |
| per-task per-model accuracy and five-bin ECE, the benchmark-to-downstream Spearman rank |
| correlations (with exact permutation p-values), the pre-set gate split, and the gate-pass vs |
| gate-fail downstream calibration ratios with and without the leanstral-2603 outlier; it checks |
| the recomputed Spearman values and gate split against the shipped summary tcb_criterion.json. |
| |
| python3 analyze_criterion.py |
| |
| Inputs: tcb_criterion_trials.json (downstream SST-2 + battery trials) and tcb_v03_trials.json |
| (the benchmark panel under the validated protocol). Stdlib only. |
| """ |
| import json, pathlib, itertools |
| from collections import defaultdict |
|
|
| HERE = pathlib.Path(__file__).resolve().parent |
| CRIT = HERE / ".." / "data" / "tcb_criterion_trials.json" |
| BENCH = HERE / ".." / "data" / "tcb_v03_trials.json" |
| SUMMARY = HERE / ".." / "data" / "tcb_criterion.json" |
| GATE_ACC, GATE_ECE = 0.90, 0.05 |
|
|
|
|
| def ece5(items): |
| bins = defaultdict(list) |
| for cp, corr in items: |
| bins[min(4, int(cp * 5))].append((cp, corr)) |
| n = len(items) |
| return sum(len(xs) / n * abs(sum(x[0] for x in xs) / len(xs) - sum(x[1] for x in xs) / len(xs)) |
| for xs in bins.values()) |
|
|
|
|
| def metrics(trials): |
| |
| |
| |
| by = defaultdict(list) |
| for t in trials: |
| if t.get("answer") is None or t.get("correct") is None: |
| continue |
| cp = 0.5 if t.get("conf") is None else t["conf"] / 100.0 |
| by[t["model"]].append((cp, t["correct"])) |
| acc = {m: sum(c for _, c in xs) / len(xs) for m, xs in by.items()} |
| ece = {m: ece5(xs) for m, xs in by.items()} |
| return acc, ece |
|
|
|
|
| def avg_ranks(vals): |
| order = sorted(range(len(vals)), key=lambda i: vals[i]) |
| r = [0.0] * len(vals) |
| i = 0 |
| while i < len(vals): |
| j = i |
| while j + 1 < len(vals) and vals[order[j + 1]] == vals[order[i]]: |
| j += 1 |
| for k in range(i, j + 1): |
| r[order[k]] = sum(range(i, j + 1)) / (j - i + 1) |
| i = j + 1 |
| return r |
|
|
|
|
| def pearson(x, y): |
| n = len(x); mx = sum(x) / n; my = sum(y) / n |
| sxy = sum((x[i] - mx) * (y[i] - my) for i in range(n)) |
| sx = sum((x[i] - mx) ** 2 for i in range(n)) ** 0.5 |
| sy = sum((y[i] - my) ** 2 for i in range(n)) ** 0.5 |
| return sxy / (sx * sy) if sx * sy else 0.0 |
|
|
|
|
| def spearman(xv, yv): |
| return pearson(avg_ranks(xv), avg_ranks(yv)) |
|
|
|
|
| def perm_p(xv, yv): |
| obs = spearman(xv, yv); xr = avg_ranks(xv); cnt = tot = 0 |
| for perm in itertools.permutations(range(len(yv))): |
| r = pearson(xr, avg_ranks([yv[i] for i in perm])); tot += 1 |
| if abs(r) >= abs(obs) - 1e-12: |
| cnt += 1 |
| return obs, cnt / tot |
|
|
|
|
| def main(): |
| crit = json.load(open(CRIT)) |
| bench = json.load(open(BENCH)) |
| summary = json.load(open(SUMMARY)) |
| bacc, bece = metrics(bench) |
| gate_pass = set(summary["gate"]["pass"]) |
| mismatches = 0 |
| for task in ("sst2", "battery"): |
| tr = [t for t in crit if t["task"] == task] |
| acc, ece = metrics(tr) |
| ms = [m for m in bacc if m in acc] |
| ra, pa = perm_p([bacc[m] for m in ms], [acc[m] for m in ms]) |
| re, pe = perm_p([bece[m] for m in ms], [ece[m] for m in ms]) |
| passm = [m for m in ms if m in gate_pass] |
| failm = [m for m in ms if m not in gate_pass] |
| fail_no = [m for m in failm if "leanstral" not in m] |
| mean = lambda d, g: sum(d[m] for m in g) / len(g) |
| pf, ff = mean(ece, passm), mean(ece, failm) |
| ff2 = mean(ece, fail_no) |
| print(f"--- {task}: n_models={len(ms)}") |
| print(f" accuracy rank: rho={ra:.3f} permutation p={pa:.3f}") |
| print(f" calib (ECE) rank: rho={re:.3f} permutation p={pe:.3f}") |
| print(f" gate-pass ECE={pf:.3f} gate-fail ECE={ff:.3f} ratio={ff/pf:.1f}x" |
| f" | without leanstral fail ECE={ff2:.3f} ratio={ff2/pf:.1f}x") |
| ref = summary["per_task"][task] |
| for label, got, want in [("spearman_acc", ra, ref["spearman_tcbacc_vs_acc"]), |
| ("spearman_ece", re, ref["spearman_tcbece_vs_ece"])]: |
| if abs(got - want) > 5e-3: |
| print(f" MISMATCH {label}: recomputed {got:.3f} vs shipped {want:.3f}") |
| mismatches += 1 |
| print("CRITERION OK" if mismatches == 0 else f"{mismatches} MISMATCHES") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
