Update app.py

app.py

@@ -106,15 +106,16 @@ THRESHOLDS = {
     "nonabusive": 1.0
 }
 
+
 PATTERN_WEIGHTS = {
     "recovery": 0.7,
     "control": 1.4,
-    "gaslighting": 1.50,
+    "gaslighting": 1.3,
     "guilt tripping": 1.2,
     "dismissiveness": 0.9,
-    "blame shifting": 0.8,
+    "blame shifting": 1.0,  # Increased from 0.8
     "projection": 0.5,
-    "insults": 1.4,
+    "insults": 1.4,        # Reduced from 2.1
     "contradictory statements": 1.0,
     "obscure language": 0.9,
     "nonabusive": 0.0
@@ -331,50 +332,51 @@ def compute_abuse_score(matched_scores, sentiment):
     """Compute abuse score from matched patterns and sentiment"""
     try:
         if not matched_scores:
+            logger.debug("No matched scores, returning 0")
             return 0.0
             
+        # Calculate weighted score
         total_weight = sum(weight for _, _, weight in matched_scores)
         if total_weight == 0:
+            logger.debug("Total weight is 0, returning 0")
             return 0.0
         
+        # Get highest pattern scores
         pattern_scores = [(label, score) for label, score, _ in matched_scores]
         sorted_scores = sorted(pattern_scores, key=lambda x: x[1], reverse=True)
+        logger.debug(f"Sorted pattern scores: {sorted_scores}")
         
+        # Base score calculation
         weighted_sum = sum(score * weight for _, score, weight in matched_scores)
         base_score = (weighted_sum / total_weight) * 100
-        
-        # Pattern combination multipliers
-        if len(matched_scores) >= 3:
-            base_score *= 1.2
-        
-        high_severity_patterns = {'gaslighting', 'control', 'blame shifting'}
-        if any(label in high_severity_patterns for label, _, _ in matched_scores):
-            base_score *= 1.15
-        
-        if any(score > 0.6 for _, score, _ in matched_scores):
-            base_score *= 1.1
+        logger.debug(f"Initial base score: {base_score:.1f}")
+        
+        # Cap maximum score based on pattern severity
+        max_score = 85.0  # Set maximum possible score
+        if any(label in {'control', 'gaslighting'} for label, _, _ in matched_scores):
+            max_score = 90.0
+            logger.debug(f"Increased max score to {max_score} due to high severity patterns")
+        
+        # Apply diminishing returns for multiple patterns
+        if len(matched_scores) > 1:
+            multiplier = 1 + (0.1 * (len(matched_scores) - 1))
+            base_score *= multiplier
+            logger.debug(f"Applied multiplier {multiplier:.2f} for {len(matched_scores)} patterns")
             
-        high_scores = len([score for _, score, _ in matched_scores if score > 0.5])
-        if high_scores >= 2:
-            base_score *= 1.15
-        
-        # Apply sentiment modifiers
+        # Apply sentiment modifier
         if sentiment == "supportive":
-            if any(label in high_severity_patterns for label, _, _ in matched_scores):
-                base_score *= 0.9
-            else:
-                base_score *= 0.85
-        elif sentiment == "undermining":
-            base_score *= 1.15
-        
-        if any(score > 0.6 for _, score, _ in matched_scores):
-            base_score = max(base_score, 65.0)
+            base_score *= 0.85
+            logger.debug("Applied 15% reduction for supportive sentiment")
         
-        return min(round(base_score, 1), 100.0)
+        final_score = min(round(base_score, 1), max_score)
+        logger.debug(f"Final abuse score: {final_score}")
+        return final_score
+
     except Exception as e:
         logger.error(f"Error computing abuse score: {e}")
         return 0.0
 
+
 @spaces.GPU
 def analyze_single_message(text, thresholds):
     """Analyze a single message for abuse patterns"""