v1.1: upgraded weights (b6472511…), re-measured metrics, refreshed docs/figures/gifs, bundled calib/, leaderboard results

calib/calibrate.py

@@ -0,0 +1,23 @@
+"""apply_gamma — per-decision power calibration for Phocinae-Largha-150M-v1.
+
+q ∝ p ** gamma  (per decision, normalized). Monotone; argmax and the 0.5
+threshold of noul answers are unchanged, so accuracy is invariant.
+Recommended: gamma=4.33 (en face), gamma=2.51 (zh face). gamma=1 = raw.
+"""
+from __future__ import annotations
+
+
+def apply_gamma(probs: dict, gamma: float = 1.0) -> dict:
+    """probs: {option_id: probability}. Returns calibrated dict (same keys)."""
+    if gamma == 1.0:
+        return dict(probs)
+    powered = {k: (max(float(v), 0.0) ** gamma) for k, v in probs.items()}
+    s = sum(powered.values())
+    if s <= 0:
+        raise ValueError("apply_gamma: all probabilities are zero")
+    return {k: v / s for k, v in powered.items()}
+
+
+def calibrate_response(probabilities: dict, gamma: float = 4.33) -> dict:
+    """Calibrate a full response: {qid: {option: prob}} -> {qid: {option: prob}}."""
+    return {qid: apply_gamma(p, gamma) for qid, p in probabilities.items()}