Update app.py

app.py

@@ -7,7 +7,7 @@ from transformers import pipeline
 # Load sentiment analysis model
 sentiment_analyzer = pipeline("sentiment-analysis")
 
-# Load model and tokenizer with trust_remote_code in case it's needed
+# Load model and tokenizer
 model_name = "SamanthaStorm/abuse-pattern-detector-v2"
 model = RobertaForSequenceClassification.from_pretrained(model_name, trust_remote_code=True)
 tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True)
@@ -21,11 +21,11 @@ LABELS = [
     "extreme_control"
 ]
 
-# Custom thresholds for each label (make sure these match your original settings)
+# Custom thresholds for each label
 THRESHOLDS = {
     "gaslighting": 0.25,
     "mockery": 0.15,
-    "dismissiveness": 0.30,  # original value, not 0.30
+    "dismissiveness": 0.30,
     "control": 0.43,
     "guilt_tripping": 0.19,
     "apology_baiting": 0.45,
@@ -41,30 +41,11 @@ THRESHOLDS = {
     "suicidal_threat": 0.45,
     "physical_threat": 0.20,
     "extreme_control": 0.36
-
 }
 
-# Define label groups using slicing (first 15: abuse patterns, last 3: danger cues)
 PATTERN_LABELS = LABELS[:15]
 DANGER_LABELS = LABELS[15:18]
 
-def calculate_abuse_level(scores, thresholds):
-    triggered_scores = [score for label, score in zip(LABELS, scores) if score > thresholds[label]]
-    if not triggered_scores:
-        return 0.0
-    return round(np.mean(triggered_scores) * 100, 2)
-
-def interpret_abuse_level(score):
-    if score > 80:
-        return "Extreme / High Risk"
-    elif score > 60:
-        return "Severe / Harmful Pattern Present"
-    elif score > 40:
-        return "Likely Abuse"
-    elif score > 20:
-        return "Mild Concern"
-    else:
-        return "Very Low / Likely Safe"
 EXPLANATIONS = {
     "gaslighting": "Gaslighting involves making someone question their own reality or perceptions, often causing them to feel confused or insecure.",
     "blame_shifting": "Blame-shifting is when one person redirects the responsibility for an issue onto someone else, avoiding accountability.",
@@ -84,63 +65,66 @@ EXPLANATIONS = {
     "manipulation": "Manipulation refers to using deceptive tactics to control or influence someone’s emotions, decisions, or behavior to serve the manipulator’s own interests.",
     "non_abusive": "Non-abusive language is communication that is respectful, empathetic, and free of harmful behaviors or manipulation."
 }
+
+def calculate_abuse_level(scores, thresholds):
+    triggered_scores = [score for label, score in zip(LABELS, scores) if score > thresholds[label]]
+    if not triggered_scores:
+        return 0.0
+    return round(np.mean(triggered_scores) * 100, 2)
+
+def interpret_abuse_level(score):
+    if score > 80:
+        return "Extreme / High Risk"
+    elif score > 60:
+        return "Severe / Harmful Pattern Present"
+    elif score > 40:
+        return "Likely Abuse"
+    elif score > 20:
+        return "Mild Concern"
+    else:
+        return "Very Low / Likely Safe"
+
 def analyze_messages(input_text):
     input_text = input_text.strip()
     if not input_text:
-        return "Please enter a message for analysis.", None
-        
-     # Sentiment analysis
-    sentiment = sentiment_analyzer(input_text)[0]  # Sentiment result
+        return "Please enter a message for analysis."
+
+    sentiment = sentiment_analyzer(input_text)[0]
     sentiment_label = sentiment['label']
     sentiment_score = sentiment['score']
-    
-     # Adjust thresholds based on sentiment
+
     adjusted_thresholds = THRESHOLDS.copy()
     if sentiment_label == "NEGATIVE":
-        # Lower thresholds for negative sentiment
-        adjusted_thresholds = {key: val * 0.8 for key, val in THRESHOLDS.items()}  # Example adjustment
-        
-    # Tokenize input and generate model predictions
+        adjusted_thresholds = {key: val * 0.8 for key, val in THRESHOLDS.items()}
+
     inputs = tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
     with torch.no_grad():
         outputs = model(**inputs)
     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 > adjusted_thresholds[label] for label, score in zip(PATTERN_LABELS, scores[:15]))
     danger_flag_count = sum(score > adjusted_thresholds[label] for label, score in zip(DANGER_LABELS, scores[15:18]))
 
-    # Check if 'non_abusive' label is triggered
     non_abusive_score = scores[LABELS.index('non_abusive')]
     if non_abusive_score > adjusted_thresholds['non_abusive']:
-        # If non-abusive threshold is met, return a non-abusive classification
         return "This message is classified as non-abusive."
 
-    # 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)
     abuse_description = interpret_abuse_level(abuse_level)
 
-    # Resource logic based on the number of danger cues
     if danger_flag_count >= 2:
         resources = "Immediate assistance recommended. Please seek professional help or contact emergency services."
     else:
         resources = "For more information on abuse patterns, consider reaching out to support groups or professional counselors."
 
-    # Get top 2 highest scoring abuse patterns (excluding 'non_abusive')
     scored_patterns = [(label, score) for label, score in zip(PATTERN_LABELS, scores[:15])]
     top_patterns = sorted(scored_patterns, key=lambda x: x[1], reverse=True)[:2]
-    top_patterns_str = "\n".join([f"• {label.replace('_', ' ').title()}" for label, _ in top_patterns])
 
-    
-    top_pattern_explanations = "\n".join([f"• {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label, 'No explanation available.')}" for label, _ in top_patterns])
+    top_pattern_explanations = "\n".join([
+        f"\u2022 {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label, 'No explanation available.')}"
+        for label, _ in top_patterns
+    ])
 
-    # Format final result
     result = (
         f"Abuse Risk Score: {abuse_level}% – {abuse_description}\n\n"
         f"Most Likely Patterns:\n{top_pattern_explanations}\n\n"
@@ -152,10 +136,8 @@ def analyze_messages(input_text):
         f"Sentiment: {sentiment_label} (Confidence: {sentiment_score*100:.2f}%)"
     )
 
-    # Return both a text summary and a JSON-like dict of scores per label
     return result
 
-# Updated Gradio Interface using new component syntax
 iface = gr.Interface(
     fn=analyze_messages,
     inputs=gr.Textbox(lines=10, placeholder="Enter message here..."),