Update app.py

app.py

@@ -382,75 +382,73 @@ THREAT_MOTIFS = [
 ]
 
 
+@spaces.GPU
 @spaces.GPU
 def compute_abuse_score(matched_scores, sentiment):
     """
-    Compute abuse score from matched patterns and sentiment, with adjustments for overinflation.
+    Compute abuse score with more conservative adjustments.
     """
     if not matched_scores:
         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
+    sorted_scores = sorted(matched_scores, key=lambda x: x[1], reverse=True)
     highest_score = sorted_scores[0][1]
     num_patterns = len(matched_scores)
+
+    # Scale down base score more aggressively if multiple patterns are present
     if num_patterns > 1:
-        highest_score *= (1 - (num_patterns - 1) * 0.1)  # Reduce by 10% for each additional pattern after the first
+        highest_score *= (1 - (num_patterns - 1) * 0.2)  # Reduce by 20% for each additional pattern
 
     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
+        'gaslighting': 1.4,  # Reduced
+        'guilt tripping': 1.3, # Reduced
+        'blame shifting': 1.2, # Reduced
+        'control': 1.3,       # Reduced
+        'insults': 1.1,       # Reduced
+        'manipulation': 1.2,
+        'love bombing': 1.2,
+        'emotional blackmail': 1.4,
+        'dismissiveness': 1.1,
+        'contradictory statements': 1.1
     }
 
     for label, score, _ in matched_scores:
         if label in critical_patterns and score > 0.5:
             base_score *= critical_patterns[label]
 
-    # Combination multipliers (reduced)
+    # Further reduce combination multipliers
     if len(matched_scores) >= 2:
-        base_score *= 1.15  # Reduced from 1.3
+        base_score *= 1.1  # Reduced
     if len(matched_scores) >= 3:
-        base_score *= 1.1  # Reduced from 1.2
+        base_score *= 1.05  # Reduced
 
-    # High confidence boost (reduced)
+    # Reduce high confidence boost
     if any(score > 0.8 for _, score, _ in matched_scores):
-        base_score *= 1.1  # Reduced from 1.2
+        base_score *= 1.05  # Reduced
 
-    # Sentiment modifier
+    # Sentiment modifier (more nuanced)
     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"
+        if any(label in manipulative_patterns for label, score, _ in matched_scores if score > 0.6): # Higher threshold
+            base_score *= 0.95  # Smaller reduction for strongly manipulative "support"
+        elif any(label in manipulative_patterns for label, score, _ in matched_scores if score > 0.4): # Moderate threshold
+            base_score *= 0.9  # Moderate reduction for manipulative "support"
         else:
-            base_score *= 0.85  # Larger reduction for genuine support
+            base_score *= 0.8  # 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
+    # Reduce minimum score and threshold for activation
+    if any(score > 0.9 for _, score, _ in matched_scores):  # Higher threshold
+        base_score = max(base_score, 75.0)  # Reduced
+    elif any(score > 0.7 for _, score, _ in matched_scores):  # Moderate threshold
+        base_score = max(base_score, 60.0)  # Reduced
 
     return min(round(base_score, 1), 100.0)
 
-
-    
-    
-
 def analyze_single_message(text, thresholds):
     print("⚡ ENTERED analyze_single_message")
     stage = 1