Update app.py

app.py

@@ -382,35 +382,72 @@ THREAT_MOTIFS = [
 ]
 
 
+@spaces.GPU
 def compute_abuse_score(matched_scores, sentiment):
+    """
+    Compute abuse score from matched patterns and sentiment, with adjustments for overinflation.
+    """
     if not matched_scores:
-        return 0
-
-    # Weighted average of passed patterns
-    weighted_total = sum(score * weight for _, score, weight in matched_scores)
-    weight_sum = sum(weight for _, _, weight in matched_scores)
-    base_score = (weighted_total / weight_sum) * 100
-
-    # Boost for pattern count
-    pattern_count = len(matched_scores)
-    scale = 1.0 + 0.25 * max(0, pattern_count - 1)  # 1.25x for 2, 1.5x for 3+
-    scaled_score = base_score * scale
-
-    # Pattern floors
-    FLOORS = {
-        "control": 40,
-        "gaslighting": 30,
-        "insults": 25,
-        "aggression": 40
+        return 0.0
+        
+    # Sort matched scores by score in descending order
+    sorted_scores = sorted(matched_scores, key=lambda x: x[1], reverse=True)
+
+    # Use the highest score as the base, but scale it down if multiple patterns are present
+    highest_score = sorted_scores[0][1]
+    num_patterns = len(matched_scores)
+    if num_patterns > 1:
+        highest_score *= (1 - (num_patterns - 1) * 0.1)  # Reduce by 10% for each additional pattern after the first
+
+    base_score = highest_score * 100
+
+    # Critical patterns with adjusted weights
+    critical_patterns = {
+        'gaslighting': 1.5,
+        'guilt tripping': 1.4,
+        'blame shifting': 1.3,
+        'control': 1.4,
+        'insults': 1.2,
+        'manipulation': 1.3,  # If you have a 'manipulation' label
+        'love bombing': 1.3,
+        'emotional blackmail': 1.5,
+        'dismissiveness': 1.2,
+        'contradictory statements': 1.2
     }
-    floor = max(FLOORS.get(label, 0) for label, _, _ in matched_scores)
-    adjusted_score = max(scaled_score, floor)
 
-    # Sentiment tweak
-    if sentiment == "undermining" and adjusted_score < 50:
-        adjusted_score += 10
+    for label, score, _ in matched_scores:
+        if label in critical_patterns and score > 0.5:
+            base_score *= critical_patterns[label]
+
+    # Combination multipliers (reduced)
+    if len(matched_scores) >= 2:
+        base_score *= 1.15  # Reduced from 1.3
+    if len(matched_scores) >= 3:
+        base_score *= 1.1  # Reduced from 1.2
+
+    # High confidence boost (reduced)
+    if any(score > 0.8 for _, score, _ in matched_scores):
+        base_score *= 1.1  # Reduced from 1.2
+
+    # Sentiment modifier
+    if sentiment == "supportive":
+        manipulative_patterns = {'guilt tripping', 'gaslighting', 'blame shifting', 'love bombing'}
+        if any(label in manipulative_patterns for label, score, _ in matched_scores if score > 0.5):
+            base_score *= 0.95  # Slightly less reduction for manipulative "support"
+        else:
+            base_score *= 0.85  # Larger reduction for genuine support
+    elif sentiment == "undermining":
+        base_score *= 1.15
+
+    # Minimum score for high-confidence patterns (adjusted)
+    if any(score > 0.8 for _, score, _ in matched_scores):
+        base_score = max(base_score, 70.0)  # Reduced from 80
+    elif any(score > 0.6 for _, score, _ in matched_scores):
+        base_score = max(base_score, 55.0)  # Reduced from 65
+
+    return min(round(base_score, 1), 100.0)
+
 
-    return min(adjusted_score, 100)