Update app.py

app.py

@@ -1,23 +1,65 @@
 import gradio as gr
 import torch
 import numpy as np
-from transformers import AutoModelForSequenceClassification, AutoTokenizer, RobertaForSequenceClassification, RobertaTokenizer
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+from transformers import RobertaForSequenceClassification, RobertaTokenizer
 from motif_tagging import detect_motifs
+from abuse_type_mapping import determine_abuse_type
 
-# Load models
+# custom fine-tuned sentiment model
 sentiment_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-sentiment")
 sentiment_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-sentiment")
+
+# Load abuse pattern model
 model_name = "SamanthaStorm/autotrain-c1un8-p8vzo"
 model = RobertaForSequenceClassification.from_pretrained(model_name, trust_remote_code=True)
 tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True)
 
-LABELS = [...]
-THRESHOLDS = {...}
+LABELS = [
+    "gaslighting", "mockery", "dismissiveness", "control", "guilt_tripping", "apology_baiting", "blame_shifting", "projection",
+    "contradictory_statements", "manipulation", "deflection", "insults", "obscure_formal", "recovery_phase", "non_abusive",
+    "suicidal_threat", "physical_threat", "extreme_control"
+]
+
+THRESHOLDS = {
+    "gaslighting": 0.25, "mockery": 0.15, "dismissiveness": 0.45, "control": 0.43, "guilt_tripping": 0.15,
+    "apology_baiting": 0.2, "blame_shifting": 0.23, "projection": 0.50, "contradictory_statements": 0.25,
+    "manipulation": 0.25, "deflection": 0.30, "insults": 0.34, "obscure_formal": 0.25, "recovery_phase": 0.25,
+    "non_abusive": 2.0, "suicidal_threat": 0.45, "physical_threat": 0.02, "extreme_control": 0.30
+}
+
 PATTERN_LABELS = LABELS[:15]
 DANGER_LABELS = LABELS[15:18]
-EXPLANATIONS = {...}
-PATTERN_WEIGHTS = {...}
 
+EXPLANATIONS = {
+    "gaslighting": "Gaslighting involves making someone question their own reality or perceptions...",
+    "blame_shifting": "Blame-shifting is when one person redirects the responsibility...",
+    "projection": "Projection involves accusing the victim of behaviors the abuser exhibits.",
+    "dismissiveness": "Dismissiveness is belittling or disregarding another person’s feelings.",
+    "mockery": "Mockery ridicules someone in a hurtful, humiliating way.",
+    "recovery_phase": "Recovery phase dismisses someone's emotional healing process.",
+    "insults": "Insults are derogatory remarks aimed at degrading someone.",
+    "apology_baiting": "Apology-baiting manipulates victims into apologizing for abuser's behavior.",
+    "deflection": "Deflection avoids accountability by redirecting blame.",
+    "control": "Control restricts autonomy through manipulation or coercion.",
+    "extreme_control": "Extreme control dominates decisions and behaviors entirely.",
+    "physical_threat": "Physical threats signal risk of bodily harm.",
+    "suicidal_threat": "Suicidal threats manipulate others using self-harm threats.",
+    "guilt_tripping": "Guilt-tripping uses guilt to manipulate someone’s actions.",
+    "manipulation": "Manipulation deceives to influence or control outcomes.",
+    "non_abusive": "Non-abusive language is respectful and free of coercion.",
+    "obscure_formal": "Obscure/formal language manipulates through confusion or superiority."
+}
+
+PATTERN_WEIGHTS = {
+    "physical_threat": 1.5,
+    "suicidal_threat": 1.4,
+    "extreme_control": 1.5,
+    "gaslighting": 1.3,
+    "control": 1.2,
+    "dismissiveness": 0.8,
+    "non_abusive": 0.0
+}
 
 def custom_sentiment(text):
     inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
@@ -30,7 +72,6 @@ def custom_sentiment(text):
     score = probs[0][label_idx].item()
     return {"label": label, "score": score}
 
-
 def calculate_abuse_level(scores, thresholds, motif_hits=None):
     weighted_scores = []
     for label, score in zip(LABELS, scores):
@@ -39,11 +80,10 @@ def calculate_abuse_level(scores, thresholds, motif_hits=None):
             weighted_scores.append(score * weight)
     base_score = round(np.mean(weighted_scores) * 100, 2) if weighted_scores else 0.0
     motif_hits = motif_hits or []
-    if any(label in motif_hits for label in DANGER_LABELS):
+    if any(label in motif_hits for label in {"physical_threat", "suicidal_threat", "extreme_control"}):
         base_score = max(base_score, 75.0)
     return base_score
 
-
 def interpret_abuse_level(score):
     if score > 80:
         return "Extreme / High Risk"
@@ -55,50 +95,59 @@ def interpret_abuse_level(score):
         return "Mild Concern"
     return "Very Low / Likely Safe"
 
-
-def analyze_single_message(text):
-    if not text.strip():
-        return "No input provided."
-    sentiment = custom_sentiment(text)
-    thresholds = {k: v * 0.8 for k, v in THRESHOLDS.items()} if sentiment['label'] == "undermining" else THRESHOLDS.copy()
+def analyze_single_message(text, contextual_flags):
     motif_flags, matched_phrases = detect_motifs(text)
+    risk_flags = list(set(contextual_flags + motif_flags)) if contextual_flags else motif_flags
+    sentiment_result = custom_sentiment(text)
+    sentiment_label = sentiment_result["label"]
+    sentiment_score = sentiment_result["score"]
+    thresholds = {k: v * 0.8 for k, v in THRESHOLDS.items()} if sentiment_label == "undermining" else THRESHOLDS.copy()
     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()
-    abuse_score = calculate_abuse_level(scores, thresholds, [label for label, _ in matched_phrases])
-    summary = interpret_abuse_level(abuse_score)
-    return f"Abuse Risk Score: {abuse_score}% — {summary}\nSentiment: {sentiment['label']} ({sentiment['score']*100:.2f}%)"
-
-
-def analyze_composite(msg1, msg2, msg3):
-    results = [analyze_single_message(t) for t in [msg1, msg2, msg3]]
-    composite_score = np.mean([
-        float(line.split('%')[0].split()[-1]) if 'Abuse Risk Score:' in line else 0
-        for line in results
-    ])
-    final_summary = interpret_abuse_level(composite_score)
-    composite_result = f"\n\nComposite Abuse Risk Score: {composite_score:.2f}% — {final_summary}"
-    return results[0], results[1], results[2], composite_result
+    threshold_labels = [label for label, score in zip(PATTERN_LABELS, scores[:15]) if score > thresholds[label]]
+    phrase_labels = [label for label, _ in matched_phrases]
+    pattern_labels_used = list(set(threshold_labels + phrase_labels))
+    abuse_level = calculate_abuse_level(scores, thresholds, motif_hits=[label for label, _ in matched_phrases])
+    abuse_description = interpret_abuse_level(abuse_level)
+    return {
+        "text": text,
+        "score": abuse_level,
+        "summary": abuse_description,
+        "sentiment": f"{sentiment_label} ({sentiment_score*100:.2f}%)",
+        "top_labels": pattern_labels_used[:2],
+        "matched_phrases": matched_phrases,
+        "flags": contextual_flags
+    }
 
+def analyze_composite(msg1, msg2, msg3, flags):
+    results = [analyze_single_message(t, flags) for t in [msg1, msg2, msg3] if t.strip()]
+    composite_score = round(np.mean([r['score'] for r in results]), 2) if results else 0.0
+    return [
+        f"Score: {r['score']}% – {r['summary']}\nSentiment: {r['sentiment']}\nFlags: {', '.join(r['flags']) if r['flags'] else 'None'}\nLabels: {', '.join(r['top_labels'])}" for r in results
+    ] + [f"Composite Abuse Score: {composite_score}%"]
 
 iface = gr.Interface(
     fn=analyze_composite,
     inputs=[
         gr.Textbox(label="Message 1"),
         gr.Textbox(label="Message 2"),
-        gr.Textbox(label="Message 3")
+        gr.Textbox(label="Message 3"),
+        gr.CheckboxGroup(label="Contextual Flags", choices=[
+            "They've threatened harm", "They isolate me", "I’ve changed my behavior out of fear",
+            "They monitor/follow me", "I feel unsafe when alone with them"
+        ])
     ],
     outputs=[
         gr.Textbox(label="Message 1 Result"),
         gr.Textbox(label="Message 2 Result"),
         gr.Textbox(label="Message 3 Result"),
-        gr.Textbox(label="Composite Score Summary")
+        gr.Textbox(label="Composite Score")
     ],
     title="Abuse Pattern Detector (Multi-Message)",
-    live=False,
-    allow_flagging="manual"
+    flagging_mode="manual"
 )
 
 if __name__ == "__main__":
-    iface.launch()
+    iface.queue().launch()