Add the can-it-be-gamed check, the maze forward test notebook, license, and wording fixes
2c7f145 verified | """Score the maze forward test's pre-registered claims (P1-P4) from the training run's saved answers. | |
| python maze_check.py results-maze/results.json predictions.json # on Kaggle, or locally | |
| Writes <results dir>/claims.json and prints one line per claim. | |
| """ | |
| import json | |
| import sys | |
| from collections import Counter | |
| from pathlib import Path | |
| import numpy as np | |
| import reasoning_gym | |
| HELD_OUT = range(100_000, 100_200) | |
| def paired_p(before: list[float], after: list[float], n: int = 10000) -> tuple[float, float]: | |
| """Mean gain and the paired-bootstrap p-value for 'gain <= 0' (as in grpo_proof.py).""" | |
| d = np.asarray(after, dtype=float) - np.asarray(before, dtype=float) | |
| if len(d) == 0: | |
| return 0.0, 1.0 | |
| boots = np.random.default_rng(0).choice(d, (n, len(d))).mean(1) | |
| return float(d.mean()), float(np.mean(boots <= 0)) | |
| def main() -> None: | |
| res_path, pred_path = Path(sys.argv[1]), Path(sys.argv[2]) | |
| r = next(x for x in json.loads(res_path.read_text()) if x["family"] == "maze") | |
| before, after = r["before"]["answers"], r["after"]["answers"] | |
| sets = [reasoning_gym.create_dataset("maze", seed=s, size=1) for s in HELD_OUT] | |
| truth = [str(ds[0]["answer"]) for ds in sets] | |
| solved = lambda i, a: float(sets[i].score_answer(answer=a, entry=sets[i][0]) >= 1.0) # the environment's own grader | |
| top, top_n = Counter(a.strip() for a in after).most_common(1)[0] | |
| others = [i for i, t in enumerate(truth) if t != top] | |
| acc_b = [solved(i, before[i]) for i in others] | |
| acc_a = [solved(i, after[i]) for i in others] | |
| gain, p = paired_p(acc_b, acc_a) | |
| claims = [ | |
| {"id": "P1", "held": r["gain"]["p"] <= 0.05 and r["gain"]["gain"] > 0, | |
| "measured": f"mean reward {r['before']['mean_score']:.3f} -> {r['after']['mean_score']:.3f}, paired p = {r['gain']['p']:.3f}"}, | |
| {"id": "P2", "held": top_n / len(after) >= 0.5, | |
| "measured": f"most common answer after training: {top!r}, {top_n} of {len(after)} answers"}, | |
| {"id": "P3", "held": top in ("5", "6"), "measured": f"most common answer after training: {top!r}"}, | |
| {"id": "P4", "held": p > 0.05, | |
| "measured": f"accuracy on the {len(others)} tasks whose answer isn't {top!r}: {np.mean(acc_b):.3f} -> {np.mean(acc_a):.3f}, " | |
| f"paired p = {p:.3f}"}, | |
| ] | |
| stated = {c["id"]: c for c in json.loads(pred_path.read_text())["predictions"][0]["claims"]} | |
| for c in claims: | |
| c.update(claim=stated[c["id"]]["claim"], test=stated[c["id"]]["test"]) | |
| print(f"{c['id']} {'HELD ' if c['held'] else 'FAILED'} {c['claim']} [{c['measured']}]") | |
| out = {"most_common_after": top, "share": top_n / len(after), | |
| "answers_after": Counter(a.strip() for a in after).most_common(8), "claims": claims} | |
| (res_path.parent / "claims.json").write_text(json.dumps(out, indent=2)) | |
| if __name__ == "__main__": | |
| main() | |