Update app.py

app.py

@@ -23,9 +23,9 @@ THRESHOLDS = {
     "mockery": 0.15,
     "dismissiveness": 0.30,  # original value, not 0.30
     "control": 0.43,
-    "guilt_tripping": 0.24,
+    "guilt_tripping": 0.19,
     "apology_baiting": 0.45,
-    "blame_shifting": 0.31,
+    "blame_shifting": 0.23,
     "projection": 0.50,
     "contradictory_statements": 0.25,
     "manipulation": 0.25,
@@ -36,12 +36,12 @@ THRESHOLDS = {
     "suicidal_threat": 0.45,
     "physical_threat": 0.31,
     "extreme_control": 0.36,
-    "non_abusive": 0.35
+    "non_abusive": 0.40
 }
 
 # Define label groups using slicing (first 14: abuse patterns, last 3: danger cues)
-PATTERN_LABELS = LABELS[:14]
-DANGER_LABELS = LABELS[14:]
+PATTERN_LABELS = LABELS[:17]
+DANGER_LABELS = LABELS[17:]
 
 def calculate_abuse_level(scores, thresholds):
     triggered_scores = [score for label, score in zip(LABELS, scores) if score > thresholds[label]]
@@ -73,8 +73,13 @@ def analyze_messages(input_text):
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
 
     # Count the number of triggered abuse pattern and danger flags based on thresholds
-    pattern_count = sum(score > THRESHOLDS[label] for label, score in zip(PATTERN_LABELS, scores[:14]))
-    danger_flag_count = sum(score > THRESHOLDS[label] for label, score in zip(DANGER_LABELS, scores[14:]))
+    pattern_count = sum(score > THRESHOLDS[label] for label, score in zip(PATTERN_LABELS, scores[:17]))
+    danger_flag_count = sum(score > THRESHOLDS[label] for label, score in zip(DANGER_LABELS, scores[17:]))
+    # Build formatted raw score display
+score_lines = [
+    f"{label:25}: {score:.3f}" for label, score in zip(PATTERN_LABELS + DANGER_LABELS, scores)
+]
+raw_score_output = "\n".join(score_lines)
 
     # Calculate overall abuse level and interpret it
     abuse_level = calculate_abuse_level(scores, THRESHOLDS)