File size: 4,056 Bytes
a2b0c04 | 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 | """What the trained model actually learned, per family: baselines, lenient re-score and Beam's audit in one file.
python training/analysis.py # writes training/kaggle/analysis.json
Baselines are answers that need no reasoning, scored on the same 200 held-out tasks as the training proof.
"""
import json
import statistics
from pathlib import Path
import reasoning_gym
ROOT = Path(__file__).resolve().parent.parent
K = ROOT / "training" / "kaggle"
SEEDS = range(100_000, 100_200)
def baseline(family: str, make_answer) -> float:
scores = []
for seed in SEEDS:
ds = reasoning_gym.create_dataset(family, seed=seed, size=1)
scores.append(ds.score_answer(answer=make_answer(ds[0]), entry=ds[0]))
return round(statistics.mean(scores), 3)
BASELINES = {
"letter_counting": ("always answer 1", lambda e: "1"),
"propositional_logic": ("join the premises with 'and'", lambda e: " ∧ ".join(f"({p})" for p in e["metadata"]["premises"])),
"word_sorting": ("repeat the input words unsorted", lambda e: ", ".join(e["metadata"]["original_words"])),
"basic_arithmetic": ("always answer 0", lambda e: "0"),
"chain_sum": ("always answer 0", lambda e: "0"),
}
LEARNED = { # our reading of the evidence below, one line per family
"basic_arithmetic": ("skill and format", "Partly skill, partly format: read leniently, the untrained model was already right on {lenient}."),
"chain_sum": ("mostly format", "Mostly format: read leniently, the untrained model was already right on {lenient}."),
"letter_counting": ("gamed", "It answers \"1\" to everything; that alone scores {baseline:.3f}."),
"propositional_logic": ("gamed", "It joins the premises with \"and\"; that alone scores {baseline:.3f}."),
"word_sorting": ("no real gain", "Repeating the input words unsorted already scores {baseline:.3f}."),
"leg_counting": ("no change", "No reward on any attempt, so nothing to learn from; this prediction could not fail."),
}
def main() -> None:
results = {r["family"]: r for r in json.loads((K / "results" / "results.json").read_text())}
lenient = json.loads((K / "lenient" / "summary.json").read_text()) if (K / "lenient" / "summary.json").exists() else {}
audit = {a["family"]: a for a in json.loads((K / "beam_audit.json").read_text())["audits"]}
rows = []
for family, r in sorted(results.items(), key=lambda kv: list(LEARNED).index(kv[0])): # the write-up's order
name, fn = BASELINES.get(family, (None, None))
base = baseline(family, fn) if fn else None
len_acc = lenient.get(family, {}).get("lenient", {}).get("accuracy")
len_mean = lenient.get(family, {}).get("lenient", {}).get("mean")
label, text = LEARNED[family]
rows.append({
"family": family, "prediction": r["prediction"], "held": r["held"],
"reward_before": round(r["before"]["mean_score"], 3), "reward_after": round(r["after"]["mean_score"], 3),
"p": r["gain"]["p"], "accuracy_before": r["before"]["accuracy"], "accuracy_after": r["after"]["accuracy"],
"lenient_accuracy_before": len_acc, "lenient_reward_before": len_mean, "baseline": {"strategy": name, "score": base} if name else None,
"learned": label, "learned_text": text.format(baseline=base or 0, lenient=f"{(len_acc or 0) * 100:g}%"),
"beam_audit": audit.get(family, {}).get("verdict"), "beam_explanation": audit.get(family, {}).get("explanation"),
})
out = {"pre_registration": {"sha256": "d11a5928bce2f1b045b593e699370612779b6b45950cbdbee256ef6a23ccc776",
"committed_utc": "2026-10-30 16:23:45", "first_checkpoint_utc": "2026-10-30 16:25"},
"held": sum(r["held"] for r in rows), "families": len(rows), "rows": rows}
(K / "analysis.json").write_text(json.dumps(out, indent=2, ensure_ascii=False))
for r in rows:
print(f"{r['family']:20s} {r['learned']:16s} audit={r['beam_audit']:15s} baseline={r['baseline']}")
if __name__ == "__main__":
main()
|