Update app.py

app.py

@@ -112,44 +112,65 @@ def calculate_darvo_score(patterns, sentiment_before, sentiment_after, motifs_fo
     contradiction_score = 1.0 if contradiction_flag else 0.0
     return round(min(0.3 * pattern_score + 0.3 * sentiment_shift_score + 0.25 * motif_score + 0.15 * contradiction_score, 1.0), 3)
 
-def generate_risk_snippet(score, top_label):
-    level = "high" if score >= 85 else "moderate" if score >= 60 else "low"
-    title, summary, advice = RISK_SNIPPETS[level]
-    return f"\n\n{title}\n{summary} (Pattern: **{top_label}**)\n💡 {advice}"
+def generate_risk_snippet(abuse_score, top_label):
+    if abuse_score >= 85:
+        risk_level = "high"
+    elif abuse_score >= 60:
+        risk_level = "moderate"
+    else:
+        risk_level = "low"
+    title, summary, advice = RISK_SNIPPETS[risk_level]
+    return f"\n\n{title}\n{summary} (Pattern: **{str(top_label)}**)\n💡 {advice}"
 
 def analyze_single_message(text, thresholds, motif_flags):
     motif_hits, matched_phrases = detect_motifs(text)
 
-    # Sentiment
+    # Sentiment Analysis
     input_ids = sentiment_tokenizer(f"emotion: {text}", return_tensors="pt").input_ids
     with torch.no_grad():
-        sentiment_out = sentiment_model.generate(input_ids)
-    emotion = sentiment_tokenizer.decode(sentiment_out[0], skip_special_tokens=True).lower()
+        outputs = sentiment_model.generate(input_ids)
+    emotion = sentiment_tokenizer.decode(outputs[0], skip_special_tokens=True).strip().lower()
     sentiment = EMOTION_TO_SENTIMENT.get(emotion, "undermining")
     sentiment_score = 0.5 if sentiment == "undermining" else 0.0
 
-    # Adjust thresholds
+    # Raise thresholds slightly if the sentiment is supportive
     adjusted_thresholds = {
-        k: v * 0.8 if sentiment == "undermining" else v * 1.2 if sentiment == "supportive" else v
+        k: v + 0.05 if sentiment == "supportive" else v
         for k, v in thresholds.items()
     }
 
+    # Contradiction Check
     contradiction_flag = detect_contradiction(text)
-    motifs = [text for _, text in matched_phrases]
 
+    # Motifs
+    motifs = [phrase for _, phrase in matched_phrases]
+
+    # Model Prediction
     inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
     with torch.no_grad():
         outputs = model(**inputs)
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
 
-    threshold_labels = [l for l, s in zip(LABELS, scores) if s > adjusted_thresholds[l]]
-    top_patterns = sorted(zip(LABELS, scores), key=lambda x: x[1], reverse=True)[:2]
-    pattern_labels = threshold_labels + [label for label, _ in matched_phrases]
+    threshold_labels = [
+        label for label, score in zip(LABELS, scores)
+        if score > adjusted_thresholds[label]
+    ]
+    top_patterns = sorted(
+        [(label, score) for label, score in zip(LABELS, scores)],
+        key=lambda x: x[1],
+        reverse=True
+    )[:2]
 
-    abuse_score = round(np.mean([s * PATTERN_WEIGHTS.get(l, 1.0) for l, s in top_patterns]) * 100, 2)
+    pattern_labels = threshold_labels + [label for label, _ in matched_phrases]
     darvo_score = calculate_darvo_score(pattern_labels, 0.0, sentiment_score, motifs, contradiction_flag)
 
-    return abuse_score, threshold_labels, top_patterns, darvo_score, {"label": sentiment, "emotion": emotion}
+    return (
+        np.mean([score for _, score in top_patterns]) * 100,
+        threshold_labels,
+        top_patterns,
+        darvo_score,
+        {"label": sentiment, "emotion": emotion}
+    )
 
 def analyze_composite(msg1, msg2, msg3, *answers_and_none):
     responses = answers_and_none[:len(ESCALATION_QUESTIONS)]