Update grader.py

grader.py

@@ -1,7 +1,15 @@
 import numpy as np
 from sentence_transformers import SentenceTransformer, util
 
-# Load model once globally
+EPS = 0.001
+
+def clamp_score(score):
+    if score <= 0.0:
+        return EPS
+    if score >= 1.0:
+        return 1.0 - EPS
+    return score
+
 _model = None
 
 def _get_model():
@@ -11,30 +19,16 @@ def _get_model():
     return _model
 
 def grade_comment(comment: str, expected_keywords: list, expert_comment: str) -> float:
-    """
-    Returns a score in [0,1] based on:
-    - semantic similarity with expert comment (70%)
-    - keyword coverage (30%)
-    - length bonus/penalty
-    """
     if not comment:
-        return 0.0
-
-    # 1. Semantic similarity
+        return clamp_score(0.0)
     model = _get_model()
     emb_comment = model.encode(comment, convert_to_tensor=True)
     emb_expert = model.encode(expert_comment, convert_to_tensor=True)
-    sim = util.pytorch_cos_sim(emb_comment, emb_expert).item()  # in [0,1]
-
-    # 2. Keyword coverage
+    sim = util.pytorch_cos_sim(emb_comment, emb_expert).item()
     comment_lower = comment.lower()
     matched = sum(1 for kw in expected_keywords if kw in comment_lower)
     kw_score = min(1.0, matched / max(1, len(expected_keywords) // 2))
-
-    # 3. Combine (70% semantic, 30% keywords)
     combined = 0.7 * sim + 0.3 * kw_score
-
-    # 4. Length bonus/penalty
     words = comment.split()
     if len(words) >= 15:
         length_bonus = 0.1
@@ -42,36 +36,27 @@ def grade_comment(comment: str, expected_keywords: list, expert_comment: str) ->
         length_bonus = -0.2
     else:
         length_bonus = 0.0
-
-    # 5. Final score, clamped
     final = combined + length_bonus
-    return max(0.0, min(1.0, final))
-
+    # Clamp to (0,1) using EPS
+    return clamp_score(final)
 
 def grade_question(question: str) -> float:
-    """Simple heuristic for question quality."""
     words = question.split()
     if len(words) < 3:
-        return 0.0
-    # Check for question words
+        return clamp_score(0.0)
     if any(q in question.lower() for q in ["what", "how", "why", "where", "when", "does", "is"]):
-        return min(1.0, len(words) / 20)
-    return 0.2
-
+        score = min(1.0, len(words) / 20)
+        return clamp_score(score)
+    return clamp_score(0.2)
 
 def grade_fix(proposed_fix: str, expected_fix_keywords: list, hidden_test: callable) -> float:
-    """Evaluates a code fix. Hidden_test can be a function that runs unit tests."""
-    # Keyword check (simplified)
     matched = sum(1 for kw in expected_fix_keywords if kw in proposed_fix.lower())
     kw_score = min(1.0, matched / max(1, len(expected_fix_keywords) // 2))
-
-    # If we have a test function, run it
     test_score = 0.0
     if hidden_test is not None:
         try:
             test_score = hidden_test(proposed_fix)
         except Exception:
             test_score = 0.0
-
-    # Weighted: 60% test, 40% keywords
-    return 0.6 * test_score + 0.4 * kw_score
+    score = 0.6 * test_score + 0.4 * kw_score
+    return clamp_score(score)