Update app.py

app.py

@@ -2,7 +2,7 @@ import gradio as gr
 import spaces
 import torch
 import numpy as np
-from transformers import AutoModelForSequenceClassification, AutoTokenizer
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline as hf_pipeline
 import re
 import matplotlib.pyplot as plt
 import io
@@ -11,41 +11,37 @@ from datetime import datetime
 from torch.nn.functional import sigmoid
 from collections import Counter
 import logging
-from transformers import pipeline as hf_pipeline
-
-# Add this with your other model loading code
-emotion_pipeline = hf_pipeline(
-    "text-classification",
-    model="j-hartmann/emotion-english-distilroberta-base",
-    top_k=6,
-    truncation=True,
-    device=0 if torch.cuda.is_available() else -1  # GPU support
-)
-
-# Add this after imports
-device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
 # Set up logging
 logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger(__name__)
 
+# Device configuration
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+logger.info(f"Using device: {device}")
 
-
-# Model initialization with error handling
 # Model initialization
 model_name = "SamanthaStorm/tether-multilabel-v4"
 model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)
 tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
 
-
 # Sentiment model
 sentiment_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-sentiment").to(device)
 sentiment_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-sentiment", use_fast=False)
 
+# Emotion pipeline
+emotion_pipeline = hf_pipeline(
+    "text-classification",
+    model="j-hartmann/emotion-english-distilroberta-base",
+    top_k=6,
+    truncation=True,
+    device=0 if torch.cuda.is_available() else -1
+)
+
 # DARVO model
 darvo_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-darvo-regressor-v1").to(device)
 darvo_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-darvo-regressor-v1", use_fast=False)
-
+darvo_model.eval()
 
 # Constants and Labels
 LABELS = [
@@ -54,12 +50,6 @@ LABELS = [
     "contradictory statements", "obscure language"
 ]
 
-TONE_LABELS = [
-    "cold invalidation", "coercive warmth", "contradictory gaslight",
-    "deflective hostility", "emotional instability", "nonabusive",
-    "performative regret", "emotional threat", "forced accountability flip"
-]
-
 SENTIMENT_LABELS = ["undermining", "supportive"]
 
 THRESHOLDS = {
@@ -89,7 +79,6 @@ PATTERN_WEIGHTS = {
     "obscure language": 0.9,
     "nonabusive": 0.0
 }
-
 ESCALATION_QUESTIONS = [
     ("Partner has access to firearms or weapons", 4),
     ("Partner threatened to kill you", 3),
@@ -110,7 +99,6 @@ RISK_STAGE_LABELS = {
     4: "🌸 Risk Stage: Calm / Honeymoon\nThis message appears supportive but may follow prior harm, minimizing it."
 }
 
-# Threat Motifs
 THREAT_MOTIFS = [
     "i'll kill you", "i'm going to hurt you", "you're dead", "you won't survive this",
     "i'll break your face", "i'll bash your head in", "i'll snap your neck",
@@ -134,14 +122,16 @@ THREAT_MOTIFS = [
     "if you just behaved, this wouldn't happen", "this is your fault",
     "you're making me hurt you", "i warned you", "you should have listened"
 ]
+
 def get_emotion_profile(text):
+    """Get emotion profile from text"""
     emotions = emotion_pipeline(text)
     if isinstance(emotions, list) and isinstance(emotions[0], list):
         emotions = emotions[0]
     return {e['label'].lower(): round(e['score'], 3) for e in emotions}
 
 def get_emotional_tone_tag(text, sentiment, patterns, abuse_score):
-    # Get emotions first
+    """Get emotional tone tag based on emotions and patterns"""
     emotions = get_emotion_profile(text)
     
     sadness = emotions.get("sadness", 0)
@@ -151,10 +141,10 @@ def get_emotional_tone_tag(text, sentiment, patterns, abuse_score):
     anger = emotions.get("anger", 0)
     fear = emotions.get("fear", 0)
 
-      # 1. Performative Regret
+    # 1. Performative Regret
     if (
         sadness > 0.4 and
-        any(p in patterns for p in ["blame shifting", "guilt tripping", "recovery phase"]) and
+        any(p in patterns for p in ["blame shifting", "guilt tripping", "recovery"]) and
         (sentiment == "undermining" or abuse_score > 40)
     ):
         return "performative regret"
@@ -172,13 +162,14 @@ def get_emotional_tone_tag(text, sentiment, patterns, abuse_score):
         (neutral + disgust) > 0.5 and
         any(p in patterns for p in ["dismissiveness", "projection", "obscure language"]) and
         sentiment == "undermining"
-    ):return "cold invalidation"
+    ):
+        return "cold invalidation"
 
     # 4. Genuine Vulnerability
     if (
         (sadness + fear) > 0.5 and
         sentiment == "supportive" and
-        all(p in ["recovery phase"] for p in patterns)
+        all(p in ["recovery"] for p in patterns)
     ):
         return "genuine vulnerability"
 
@@ -205,59 +196,35 @@ def get_emotional_tone_tag(text, sentiment, patterns, abuse_score):
         sentiment == "undermining"
     ):
         return "toxic resignation"
-     # 8. Aggressive Dismissal
+
+    # 8. Aggressive Dismissal
     if (
         anger > 0.5 and
-        any(p in patterns for p in ["aggression", "insults", "control"]) and
+        any(p in patterns for p in ["insults", "control"]) and
         sentiment == "undermining"
     ):
         return "aggressive dismissal"
+
     # 9. Deflective Hostility
     if (
         (0.2 < anger < 0.7 or 0.2 < disgust < 0.7) and
-        any(p in patterns for p in ["deflection", "projection"]) and
-        sentiment == "undermining"
-    ):
-        return "deflective hostility"   
-    # 10. Mocking Detachment
-    if (
-        (neutral + joy) > 0.5 and
-        any(p in patterns for p in ["mockery", "insults", "projection"]) and
+        any(p in patterns for p in ["projection"]) and
         sentiment == "undermining"
     ):
-        return "mocking detachment"
-        # 11. Contradictory Gaslight
+        return "deflective hostility"
+
+    # 10. Contradictory Gaslight
     if (
         (joy + anger + sadness) > 0.5 and
         any(p in patterns for p in ["gaslighting", "contradictory statements"]) and
         sentiment == "undermining"
     ):
         return "contradictory gaslight"
-        # 12. Calculated Neutrality
-    if (
-        neutral > 0.6 and
-        any(p in patterns for p in ["obscure language", "deflection", "dismissiveness"]) and
-        sentiment == "undermining"
-    ):
-        return "calculated neutrality"
-     # 13. Forced Accountability Flip
-    if (
-        (anger + disgust) > 0.5 and
-        any(p in patterns for p in ["blame shifting", "manipulation", "projection"]) and
-        sentiment == "undermining"
-    ):
-        return "forced accountability flip"
-        # 14. Conditional Affection
-    if (
-        joy > 0.4 and
-        any(p in patterns for p in ["apology baiting", "control", "recovery phase"]) and
-        sentiment == "undermining"
-    ):
-        return "conditional affection"
-    
+
+    # 11. Forced Accountability Flip
     if (
         (anger + disgust) > 0.5 and
-        any(p in patterns for p in ["blame shifting", "projection", "deflection"]) and
+        any(p in patterns for p in ["blame shifting", "projection"]) and
         sentiment == "undermining"
     ):
         return "forced accountability flip"
@@ -268,9 +235,9 @@ def get_emotional_tone_tag(text, sentiment, patterns, abuse_score):
         sentiment == "undermining"
     ):
         return "emotional instability"
-        
-    return None
 
+    return "neutral"
+@spaces.GPU
 def predict_darvo_score(text):
     """Predict DARVO score for given text"""
     try:
@@ -305,17 +272,13 @@ def get_risk_stage(patterns, sentiment):
         logger.error(f"Error determining risk stage: {e}")
         return 1
 
-
 @spaces.GPU
 def compute_abuse_score(matched_scores, sentiment):
-    """
-    Compute abuse score from matched patterns and sentiment
-    """
+    """Compute abuse score from matched patterns and sentiment"""
     try:
         if not matched_scores:
             return 0.0
             
-        # Calculate weighted score
         total_weight = sum(weight for _, _, weight in matched_scores)
         if total_weight == 0:
             return 0.0
@@ -326,7 +289,7 @@ def compute_abuse_score(matched_scores, sentiment):
         weighted_sum = sum(score * weight for _, score, weight in matched_scores)
         base_score = (weighted_sum / total_weight) * 100
         
-        # Apply multipliers
+        # Pattern combination multipliers
         if len(matched_scores) >= 3:
             base_score *= 1.2
         
@@ -358,7 +321,6 @@ def compute_abuse_score(matched_scores, sentiment):
         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"""
@@ -374,19 +336,10 @@ def analyze_single_message(text, thresholds):
         explicit_abuse_words = ['fuck', 'bitch', 'shit', 'ass', 'dick']
         explicit_abuse = any(word in text.lower() for word in explicit_abuse_words)
         logger.debug(f"Explicit abuse detected: {explicit_abuse}")
-         # Get sentiment
-        sent_inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
-        sent_inputs = {k: v.to(device) for k, v in sent_inputs.items()}
-        with torch.no_grad():
-            sent_logits = sentiment_model(**sent_inputs).logits[0]
-        sent_probs = torch.softmax(sent_logits, dim=-1).cpu().numpy()
-        sentiment = SENTIMENT_LABELS[int(np.argmax(sent_probs))]
         
-        # Get tone using emotion-based approach
-        tone_tag = get_emotional_tone_tag(text, sentiment, threshold_labels, abuse_score)
         # Abuse model inference 
         inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
-        inputs = {k: v.to(device) for k, v in inputs.items()}  # Move to GPU
+        inputs = {k: v.to(device) for k, v in inputs.items()}
         
         with torch.no_grad():
             outputs = model(**inputs)
@@ -415,7 +368,7 @@ def analyze_single_message(text, thresholds):
             if explicit_abuse:
                 base_threshold *= 0.5
             if score > base_threshold:
-                if label not in threshold_labels:  # Avoid duplicates
+                if label not in threshold_labels:
                     threshold_labels.append(label)
         
         logger.debug("\nLabels that passed thresholds:", threshold_labels)
@@ -428,8 +381,6 @@ def analyze_single_message(text, thresholds):
             if explicit_abuse and label == "insults":
                 weight *= 1.5
             matched_scores.append((label, score, weight))
-        
-        logger.debug("\nMatched scores (label, score, weight):", matched_scores)
 
         # Get sentiment
         sent_inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
@@ -438,38 +389,23 @@ def analyze_single_message(text, thresholds):
             sent_logits = sentiment_model(**sent_inputs).logits[0]
         sent_probs = torch.softmax(sent_logits, dim=-1).cpu().numpy()
         sentiment = SENTIMENT_LABELS[int(np.argmax(sent_probs))]
-        logger.debug(f"\nDetected sentiment: {sentiment}")
-
-        # Get tone
-        tone_inputs = tone_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
-        tone_inputs = {k: v.to(device) for k, v in tone_inputs.items()}
-        with torch.no_grad():
-            tone_logits = tone_model(**tone_inputs).logits[0]
-        tone_probs = torch.sigmoid(tone_logits).cpu().numpy()
-        tone_tag = TONE_LABELS[int(np.argmax(tone_probs))]
-        logger.debug(f"Detected tone: {tone_tag}")
-
-        # Get DARVO score
-        darvo_score = predict_darvo_score(text)
-        logger.debug(f"DARVO score: {darvo_score}")
 
         # Calculate abuse score
-        if not matched_scores:
-            logger.debug("No matched scores, returning 0")
-            return 0.0, [], [], {"label": "undermining"}, 2 if explicit_abuse else 1, 0.0, None
-
         abuse_score = compute_abuse_score(matched_scores, sentiment)
-        
         if explicit_abuse:
             abuse_score = max(abuse_score, 70.0)
-        
-        logger.debug(f"\nCalculated abuse score: {abuse_score}")
+
+        # Get DARVO score
+        darvo_score = predict_darvo_score(text)
+
+        # Get tone using emotion-based approach
+        tone_tag = get_emotional_tone_tag(text, sentiment, threshold_labels, abuse_score)
 
         # Set stage
         stage = 2 if explicit_abuse or abuse_score > 70 else 1
-        logger.debug(f"Final stage: {stage}")
         
         logger.debug("=== DEBUG END ===\n")
+
         return abuse_score, threshold_labels, matched_scores, {"label": sentiment}, stage, darvo_score, tone_tag
 
     except Exception as e:
@@ -547,7 +483,6 @@ def analyze_composite(msg1, msg2, msg3, *answers_and_none):
         responses_checked = any(answers_and_none[:-1])
         none_selected = not responses_checked and none_selected_checked
 
-        # Determine escalation score
         if none_selected:
             escalation_score = 0
             escalation_note = "Checklist completed: no danger items reported."
@@ -599,7 +534,7 @@ def analyze_composite(msg1, msg2, msg3, *answers_and_none):
         high = {'control'}
         moderate = {'gaslighting', 'dismissiveness', 'obscure language', 'insults', 
                    'contradictory statements', 'guilt tripping'}
-        low = {'blame shifting', 'projection', 'recovery phase'}
+        low = {'blame shifting', 'projection', 'recovery'}
         
         counts = {'high': 0, 'moderate': 0, 'low': 0}
         for label in predicted_labels:
@@ -610,7 +545,7 @@ def analyze_composite(msg1, msg2, msg3, *answers_and_none):
             elif label in low:
                 counts['low'] += 1
 
-        # Determine pattern escalation risk
+        # Pattern escalation logic
         if counts['high'] >= 2 and counts['moderate'] >= 2:
             pattern_escalation_risk = "Critical"
         elif (counts['high'] >= 2 and counts['moderate'] >= 1) or \
@@ -683,40 +618,63 @@ def analyze_composite(msg1, msg2, msg3, *answers_and_none):
                 f"• Escalation Bump: +{escalation_bump} (from DARVO, tone, intensity, etc.)"
             )
 
-        # Calculate composite abuse score
+        # Composite Abuse Score
         composite_abuse = int(round(sum(abuse_scores) / len(abuse_scores)))
 
         # Get most common stage
         most_common_stage = max(set(stages), key=stages.count)
         stage_text = RISK_STAGE_LABELS[most_common_stage]
 
-        # Get top labels
-        top_labels = []
-        for result, _ in results:
-            threshold_labels = result[1]
-            if threshold_labels:
-                top_labels.append(threshold_labels[0])
-            else:
-                top_labels.append("none")
+        # Build output text
+        out = f"Abuse Intensity: {composite_abuse}%\n"
+        out += "📊 This reflects the strength and severity of detected abuse patterns in the message(s).\n\n"
 
-        # Calculate average DARVO score
+        # Add risk assessment
+        risk_level = (
+            "Critical" if composite_abuse >= 85 or hybrid_score >= 20 else
+            "High" if composite_abuse >= 70 or hybrid_score >= 15 else
+            "Moderate" if composite_abuse >= 50 or hybrid_score >= 10 else
+            "Low"
+        )
+        
+        risk_descriptions = {
+            "Critical": (
+                "🚨 **Risk Level: Critical**\n"
+                "Multiple severe abuse patterns detected. This situation shows signs of "
+                "dangerous escalation and immediate intervention may be needed."
+            ),
+            "High": (
+                "⚠️ **Risk Level: High**\n"
+                "Strong abuse patterns detected. This situation shows concerning "
+                "signs of manipulation and control."
+            ),
+            "Moderate": (
+                "⚡ **Risk Level: Moderate**\n"
+                "Concerning patterns detected. While not severe, these behaviors "
+                "indicate unhealthy relationship dynamics."
+            ),
+            "Low": (
+                "📝 **Risk Level: Low**\n"
+                "Minor concerning patterns detected. While present, the detected "
+                "behaviors are subtle or infrequent."
+            )
+        }
+        
+        out += risk_descriptions[risk_level]
+        out += f"\n\n{stage_text}"
+
+        # Add DARVO analysis
         avg_darvo = round(sum(darvo_scores) / len(darvo_scores), 3)
-        darvo_blurb = ""
         if avg_darvo > 0.25:
             level = "moderate" if avg_darvo < 0.65 else "high"
-            darvo_blurb = f"\n\n🎭 **DARVO Score: {avg_darvo}** → This indicates a **{level} likelihood** of narrative reversal (DARVO), where the speaker may be denying, attacking, or reversing blame."
+            out += f"\n\n🎭 **DARVO Score: {avg_darvo}** → This indicates a **{level} likelihood** of narrative reversal (DARVO), where the speaker may be denying, attacking, or reversing blame."
 
-        # Build output text
-        out = f"Abuse Intensity: {composite_abuse}%\n"
-        out += "📊 This reflects the strength and severity of detected abuse patterns in the message(s).\n\n"
-        out += generate_risk_snippet(composite_abuse, top_labels[0], hybrid_score, most_common_stage)
-        out += f"\n\n{stage_text}"
-        out += darvo_blurb
+        # Add emotional tones
         out += "\n\n🎭 **Emotional Tones Detected:**\n"
         for i, tone in enumerate(tone_tags):
             out += f"• Message {i+1}: *{tone or 'none'}*\n"
 
-        # Add threat section
+        # Add threats section
         if flat_threats:
             out += "\n\n🚨 **Immediate Danger Threats Detected:**\n"
             for t in set(flat_threats):
@@ -767,7 +725,7 @@ def create_interface():
         logger.error(f"Error creating interface: {e}")
         raise
 
-
+# Main execution
 if __name__ == "__main__":
     try:
         demo = create_interface()
@@ -777,4 +735,5 @@ if __name__ == "__main__":
             share=False
         )
     except Exception as e:
-        print(f"Error launching app: {e}")
+        logger.error(f"Failed to launch app: {e}")
+        raise