Update app.py

app.py

@@ -2,18 +2,15 @@ import gradio as gr
 import torch
 import numpy as np
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
-from transformers import RobertaForSequenceClassification, RobertaTokenizer
 from motif_tagging import detect_motifs
-from abuse_type_mapping import determine_abuse_type
 
-# custom fine-tuned sentiment model
+# Load models
 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)
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
 
 LABELS = [
     "gaslighting", "mockery", "dismissiveness", "control", "guilt_tripping", "apology_baiting", "blame_shifting", "projection",
@@ -28,36 +25,32 @@ THRESHOLDS = {
     "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 = {
     "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."
+    "blame_shifting": "Redirecting responsibility to the victim...",
+    "projection": "Accusing the victim of behaviors the abuser exhibits...",
+    "dismissiveness": "Belittling or disregarding someone's feelings...",
+    "mockery": "Ridiculing someone in a hurtful, humiliating way...",
+    "recovery_phase": "Dismissing someone's emotional healing...",
+    "insults": "Derogatory remarks aimed at degrading someone...",
+    "apology_baiting": "Manipulating victims into apologizing for abuse...",
+    "deflection": "Redirecting blame to avoid accountability...",
+    "control": "Restricting autonomy through manipulation...",
+    "extreme_control": "Dominating decisions and behaviors entirely...",
+    "physical_threat": "Signals risk of bodily harm...",
+    "suicidal_threat": "Manipulates others using self-harm threats...",
+    "guilt_tripping": "Uses guilt to manipulate someone's actions...",
+    "manipulation": "Deceives to influence or control outcomes...",
+    "non_abusive": "Respectful and free of coercion...",
+    "obscure_formal": "Uses confusion/superiority to manipulate..."
 }
 
+DANGER_LABELS = LABELS[15:18]
+PATTERN_LABELS = LABELS[:15]
+
 PATTERN_WEIGHTS = {
-    "physical_threat": 1.5,
-    "suicidal_threat": 1.4,
-    "extreme_control": 1.5,
-    "gaslighting": 1.3,
-    "control": 1.2,
-    "dismissiveness": 0.8,
+    "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
 }
 
@@ -67,73 +60,50 @@ def custom_sentiment(text):
         outputs = sentiment_model(**inputs)
         probs = torch.nn.functional.softmax(outputs.logits, dim=1)
         label_idx = torch.argmax(probs).item()
-    label_map = {0: "supportive", 1: "undermining"}
-    label = label_map[label_idx]
-    score = probs[0][label_idx].item()
-    return {"label": label, "score": score}
+    return {"label": "supportive" if label_idx == 0 else "undermining", "score": probs[0][label_idx].item()}
 
 def calculate_abuse_level(scores, thresholds, motif_hits=None):
-    weighted_scores = []
-    for label, score in zip(LABELS, scores):
-        if score > thresholds[label]:
-            weight = PATTERN_WEIGHTS.get(label, 1.0)
-            weighted_scores.append(score * weight)
+    weighted_scores = [score * PATTERN_WEIGHTS.get(label, 1.0) for label, score in zip(LABELS, scores) if score > thresholds[label]]
     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 {"physical_threat", "suicidal_threat", "extreme_control"}):
+    if any(label in (motif_hits or []) for label in DANGER_LABELS):
         base_score = max(base_score, 75.0)
     return base_score
 
 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"
+    if score > 80: return "Extreme / High Risk"
+    if score > 60: return "Severe / Harmful Pattern Present"
+    if score > 40: return "Likely Abuse"
+    if score > 20: return "Mild Concern"
     return "Very Low / Likely Safe"
 
-def analyze_single_message(text, contextual_flags):
+def analyze_single_message(text, thresholds, context_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()
+    sentiment = custom_sentiment(text)
+    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()
-    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])
+        scores = torch.sigmoid(model(**inputs).logits.squeeze(0)).numpy()
+    labels_used = list(set([l for l, s in zip(PATTERN_LABELS, scores[:15]) if s > thresholds[l]] + [l for l, _ in matched_phrases]))
+    abuse_level = calculate_abuse_level(scores, thresholds, motif_hits=[l for l, _ 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
-    }
+    danger_count = sum(scores[LABELS.index(lbl)] > thresholds[lbl] for lbl in DANGER_LABELS)
+    output = f"Score: {abuse_level}% – {abuse_description}\nLabels: {', '.join(labels_used)}"
+    return output, abuse_level
 
 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}%"]
+    thresholds = THRESHOLDS.copy()
+    results = [analyze_single_message(t, thresholds, flags) for t in [msg1, msg2, msg3] if t.strip()]
+    result_texts = [r[0] for r in results]
+    composite_score = round(np.mean([r[1] for r in results]), 2) if results else 0.0
+    result_texts.append(f"\nComposite Abuse Score: {composite_score}%")
+    return tuple(result_texts)
 
 iface = gr.Interface(
     fn=analyze_composite,
     inputs=[
-        gr.Textbox(label="Message 1"),
-        gr.Textbox(label="Message 2"),
-        gr.Textbox(label="Message 3"),
+        gr.Textbox(lines=3, label="Message 1"),
+        gr.Textbox(lines=3, label="Message 2"),
+        gr.Textbox(lines=3, 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"
@@ -150,4 +120,4 @@ iface = gr.Interface(
 )
 
 if __name__ == "__main__":
-    iface.queue().launch()
+    iface.launch()