Upload server/reward.py with huggingface_hub

server/reward.py

@@ -0,0 +1,63 @@
+from __future__ import annotations
+import numpy as np
+from sentence_transformers import SentenceTransformer
+
+
+def _cosine(a: np.ndarray, b: np.ndarray) -> float:
+    na = np.linalg.norm(a)
+    nb = np.linalg.norm(b)
+    if na == 0 or nb == 0:
+        return 0.0
+    return float(np.dot(a, b) / (na * nb))
+
+_embedder: SentenceTransformer | None = None
+
+
+def get_embedder() -> SentenceTransformer:
+    global _embedder
+    if _embedder is None:
+        _embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+    return _embedder
+
+
+def compute_semantic_similarity(hypothesis: str, true_secret: str) -> float:
+    embedder = get_embedder()
+    emb_h = embedder.encode([hypothesis])[0]
+    emb_t = embedder.encode([true_secret])[0]
+    raw = _cosine(emb_h, emb_t)
+    # random sentence pairs score ~0.2, normalize to [0, 1]
+    return max(0.0, (raw - 0.2) / 0.8)
+
+
+def compute_reward(
+    hypothesis: str,
+    true_secret: str,
+    n_questions_used: int,
+    max_questions: int,
+    category_predicted: str | None,
+    category_true: str,
+    hint_keywords: list[str],
+) -> dict:
+    semantic = compute_semantic_similarity(hypothesis, true_secret)
+
+    # fewer questions = higher bonus (range: 0.6 – 1.0)
+    efficiency = 1.0 - (n_questions_used / max_questions) * 0.4
+
+    category_bonus = 0.1 if category_predicted == category_true else 0.0
+
+    hypothesis_lower = hypothesis.lower()
+    kw_hits = sum(1 for kw in hint_keywords if kw.lower() in hypothesis_lower)
+    keyword_bonus = min(0.1, 0.033 * kw_hits)
+
+    base = semantic * efficiency
+    total = min(1.0, base + category_bonus + keyword_bonus)
+
+    return {
+        "reward": round(total, 4),
+        "components": {
+            "semantic": round(semantic, 4),
+            "efficiency": round(efficiency, 4),
+            "category_bonus": category_bonus,
+            "keyword_bonus": round(keyword_bonus, 4),
+        },
+    }