Update app.py

app.py

@@ -2,15 +2,18 @@ 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
 
-# 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 = AutoModelForSequenceClassification.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = RobertaForSequenceClassification.from_pretrained(model_name, trust_remote_code=True)
+tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True)
 
 LABELS = [
     "gaslighting", "mockery", "dismissiveness", "control", "guilt_tripping", "apology_baiting", "blame_shifting", "projection",
@@ -25,32 +28,36 @@ 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": "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..."
+    "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."
 }
 
-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
 }
 
@@ -60,64 +67,77 @@ def custom_sentiment(text):
         outputs = sentiment_model(**inputs)
         probs = torch.nn.functional.softmax(outputs.logits, dim=1)
         label_idx = torch.argmax(probs).item()
-    return {"label": "supportive" if label_idx == 0 else "undermining", "score": probs[0][label_idx].item()}
+    label_map = {0: "supportive", 1: "undermining"}
+    return {"label": label_map[label_idx], "score": probs[0][label_idx].item()}
 
 def calculate_abuse_level(scores, thresholds, motif_hits=None):
     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
-    if any(label in (motif_hits or []) for label in DANGER_LABELS):
+    motif_hits = motif_hits or []
+    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"
-    if score > 60: return "Severe / Harmful Pattern Present"
-    if score > 40: return "Likely Abuse"
-    if score > 20: return "Mild Concern"
+    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"
     return "Very Low / Likely Safe"
 
-def analyze_single_message(text, thresholds, context_flags):
-    motif_flags, matched_phrases = detect_motifs(text)
+def analyze_single_message(text, thresholds, motif_flags):
+    motif_hits, matched_phrases = detect_motifs(text)
     sentiment = custom_sentiment(text)
-    thresholds = {k: v * 0.8 for k, v in thresholds.items()} if sentiment['label'] == "undermining" else thresholds.copy()
+    adjusted_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():
-        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])
+        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 > adjusted_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, adjusted_thresholds, motif_hits)
     abuse_description = interpret_abuse_level(abuse_level)
-    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
+    top_patterns = sorted([(label, score) for label, score in zip(PATTERN_LABELS, scores[:15]) if label != "non_abusive"], key=lambda x: x[1], reverse=True)[:2]
+    pattern_expl = "\n".join([f"• {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label)}" for label, _ in top_patterns])
+    return abuse_level, abuse_description, pattern_expl
 
 def analyze_composite(msg1, msg2, msg3, flags):
-    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)
+    thresholds = THRESHOLDS
+    results = [analyze_single_message(m, thresholds, flags) for m in [msg1, msg2, msg3] if m.strip()]
+    if not results:
+        return "Please enter at least one message."
+    result_lines = []
+    total_score = 0
+    for i, (score, desc, patterns) in enumerate(results, 1):
+        total_score += score
+        result_lines.append(f"Message {i}: {score:.2f}% – {desc}\n{patterns}\n")
+    composite = round(total_score / len(results), 2)
+    result_lines.append(f"\nComposite Abuse Score: {composite}%")
+    return "\n\n".join(result_lines)
+
+txt_inputs = [
+    gr.Textbox(label="Message 1"),
+    gr.Textbox(label="Message 2"),
+    gr.Textbox(label="Message 3")
+]
+
+checkboxes = 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"
+])
 
 iface = gr.Interface(
     fn=analyze_composite,
-    inputs=[
-        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"
-        ])
-    ],
-    outputs=[
-        gr.Textbox(label="Message 1 Result"),
-        gr.Textbox(label="Message 2 Result"),
-        gr.Textbox(label="Message 3 Result"),
-        gr.Textbox(label="Composite Score")
-    ],
+    inputs=txt_inputs + [checkboxes],
+    outputs=gr.Textbox(label="Results"),
     title="Abuse Pattern Detector (Multi-Message)",
-    flagging_mode="manual"
+    allow_flagging="manual"
 )
 
 if __name__ == "__main__":
-    iface.launch()
+    iface.launch()