Update app.py

app.py

@@ -9,25 +9,18 @@ import io
 from PIL import Image
 from datetime import datetime
 
-# ——— Constants & Utilities ———
+# ——— Constants ———
 LABELS = [
     "blame shifting", "contradictory statements", "control", "dismissiveness",
     "gaslighting", "guilt tripping", "insults", "obscure language",
     "projection", "recovery phase", "threat"
 ]
 
+# <- Restore your exact thresholds here:
 THRESHOLDS = {
-    "blame shifting": 0.3,
-    "contradictory statements": 0.3,
-    "control": 0.35,
-    "dismissiveness": 0.4,
-    "gaslighting": 0.3,
-    "guilt tripping": 0.3,
-    "insults": 0.3,
-    "obscure language": 0.4,
-    "projection": 0.4,
-    "recovery phase": 0.35,
-    "threat": 0.3
+    "blame shifting": 0.3, "contradictory statements": 0.3, "control": 0.35, "dismissiveness": 0.4,
+    "gaslighting": 0.3, "guilt tripping": 0.3, "insults": 0.3, "obscure language": 0.4,
+    "projection": 0.4, "recovery phase": 0.35, "threat": 0.3
 }
 
 PATTERN_WEIGHTS = {
@@ -39,10 +32,9 @@ PATTERN_WEIGHTS = {
     "threat": 1.5
 }
 
-DARVO_PATTERNS = set([
+DARVO_PATTERNS = {
     "blame shifting", "projection", "dismissiveness", "guilt tripping", "contradictory statements"
-])
-
+}
 DARVO_MOTIFS = [
     "I never said that.", "You’re imagining things.", "That never happened.",
     "You’re making a big deal out of nothing.", "It was just a joke.", "You’re too sensitive.",
@@ -65,15 +57,16 @@ DARVO_MOTIFS = [
     "You’re making me feel like a terrible person.", "You’re always blaming me for everything.",
     "You’re the one who’s abusive.", "You’re the one who’s controlling.", "You’re the one who’s manipulative.",
     "You’re the one who’s toxic.", "You’re the one who’s gaslighting me.",
-    "You’re the one who’s always putting me down.",
+    "You’re the one who’s always putting me down.", "You’re the one who’s always making me feel bad.",
+    "You’re the one who’s always making me feel like I’m not good enough.",
     "You’re the one who’s always making me feel like I’m the problem.",
-    "You’re the one who’s always making me feel like the bad guy.",
-    "You’re the one who’s always making me feel like the villain.",
-    "You’re the one who’s always making me feel like the one who needs to change.",
-    "You’re the one who’s always making me feel like the one who’s wrong.",
-    "You’re the one who’s always making me feel like the one who’s crazy.",
-    "You’re the one who’s always making me feel like the one who’s abusive.",
-    "You’re the one who’s always making me feel like the one who’s toxic."
+    "You’re the one who’s always making me feel like I’m the bad guy.",
+    "You’re the one who’s always making me feel like I’m the villain.",
+    "You’re the one who’s always making me feel like I’m the one who needs to change.",
+    "You’re the one who’s always making me feel like I’m the one who’s wrong.",
+    "You’re the one who’s always making me feel like I’m the one who’s crazy.",
+    "You’re the one who’s always making me feel like I’m the one who’s abusive.",
+    "You’re the one who’s always making me feel like I’m the one who’s toxic."
 ]
 
 RISK_STAGE_LABELS = {
@@ -93,9 +86,10 @@ ESCALATION_QUESTIONS = [
     ("Partner forced or coerced you into unwanted sexual acts", 3),
     ("Partner threatened to take away your children", 2),
     ("Violence has increased in frequency or severity", 3),
-    ("Partner monitors your calls/GPS/social media", 2)
+    ("Partner monitors your calls, GPS, or social media", 2)
 ]
 
+# ——— Helper Functions ———
 def detect_contradiction(message):
     patterns = [
         (r"\b(i love you).{0,15}(i hate you|you ruin everything)", re.IGNORECASE),
@@ -120,23 +114,23 @@ def calculate_darvo_score(patterns, sentiment_before, sentiment_after, motifs_fo
 def generate_risk_snippet(abuse_score, top_label, escalation_score, stage):
     label = top_label.split(" – ")[0]
     why = {
-        "control": "This message may reflect efforts to restrict someone’s autonomy.",
-        "gaslighting": "This message could be manipulating perception.",
-        "dismissiveness": "This message may include belittling, invalidating, or ignoring the other person’s experience.",
-        "insults": "Direct insults can erode emotional safety.",
-        "threat": "This message includes threatening language—a strong predictor of harm.",
-        "blame shifting": "This message may redirect responsibility to avoid accountability.",
-        "guilt tripping": "This message may induce guilt to control behavior.",
-        "recovery phase": "This message may be a tension-reset without change.",
-        "projection": "This message may attribute the speaker’s faults to the target.",
-    }.get(label, "This message contains language patterns that may affect safety.")
+        "control": "efforts to restrict autonomy.",
+        "gaslighting": "manipulating perception.",
+        "dismissiveness": "invalidating experience.",
+        "insults": "direct insults erode safety.",
+        "threat": "threatening language predicts harm.",
+        "blame shifting": "avoiding accountability.",
+        "guilt tripping": "inducing guilt to control behavior.",
+        "recovery phase": "tension-reset without change.",
+        "projection": "attributing faults to the other person."
+    }.get(label, "This message contains concerning patterns.")
     if abuse_score>=85 or escalation_score>=16:
         lvl = "high"
     elif abuse_score>=60 or escalation_score>=8:
         lvl = "moderate"
     else:
         lvl = "low"
-    return f"\n\n🛑 Risk Level: {lvl.capitalize()}\nThis message shows **{label}**.\n💡 *Why:* {why}\n"
+    return f"\n\n🛑 Risk Level: {lvl.capitalize()}\nThis message shows **{label}**.\n💡 Why: {why}\n"
 
 def detect_weapon_language(text):
     kws = ["knife","gun","bomb","kill you","shoot","explode"]
@@ -150,131 +144,120 @@ def get_risk_stage(patterns, sentiment):
         return 1
     if "recovery phase" in patterns:
         return 3
-    if sentiment == "supportive" and any(p in patterns for p in ["projection","dismissiveness"]):
+    if sentiment=="supportive" and any(p in patterns for p in ["projection","dismissiveness"]):
         return 4
     return 1
 
-# --- Visualization ---
-def generate_abuse_score_chart(dates, scores, labels):
+def generate_abuse_score_chart(dates,scores,labels):
     try:
-        parsed = [datetime.strptime(d, "%Y-%m-%d") for d in dates]
+        parsed=[datetime.strptime(d,"%Y-%m-%d") for d in dates]
     except:
-        parsed = list(range(len(dates)))
-    fig, ax = plt.subplots(figsize=(8,3))
-    ax.plot(parsed, scores, marker='o', linestyle='-', color='darkred', linewidth=2)
+        parsed=range(len(dates))
+    fig,ax=plt.subplots(figsize=(8,3))
+    ax.plot(parsed,scores,marker='o',linestyle='-',color='darkred',linewidth=2)
     for i,(x,y) in enumerate(zip(parsed,scores)):
-        ax.text(x, y+2, f"{labels[i]}\n{int(y)}%", ha='center', fontsize=8)
-    ax.set_title("Abuse Intensity Over Time")
-    ax.set_xlabel("Date"); ax.set_ylabel("Abuse Score (%)")
-    ax.set_ylim(0,105); ax.grid(True); plt.tight_layout()
-    buf = io.BytesIO(); plt.savefig(buf, format='png'); buf.seek(0)
+        ax.text(x,y+2,f"{labels[i]}\n{int(y)}%",ha='center',fontsize=8)
+    ax.set(title="Abuse Intensity Over Time",xlabel="Date",ylabel="Abuse Score (%)")
+    ax.set_ylim(0,105);ax.grid(True);plt.tight_layout()
+    buf=io.BytesIO();plt.savefig(buf,format='png');buf.seek(0)
     return Image.open(buf)
 
-# --- Load Models & Pipelines ---
-model_name = "SamanthaStorm/tether-multilabel-v2"
-model = AutoModelForSequenceClassification.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
-healthy_detector = pipeline("text-classification", model="distilbert-base-uncased-finetuned-sst-2-english")
+# ——— Load Models & Pipelines ———
+model_name="SamanthaStorm/tether-multilabel-v2"
+model=AutoModelForSequenceClassification.from_pretrained(model_name)
+tokenizer=AutoTokenizer.from_pretrained(model_name, use_fast=False)
+healthy_detector=pipeline("text-classification",model="distilbert-base-uncased-finetuned-sst-2-english")
 sst_pipeline = pipeline("sentiment-analysis", model="distilbert-base-uncased-finetuned-sst-2-english")
 
-# --- Single-Message Analysis ---
+# ——— Single-Message Analysis ———
 def analyze_single_message(text):
     # healthy bypass
-    healthy = healthy_detector(text)[0]
-    if healthy['label'] == "POSITIVE" and healthy['score'] > 0.9:
-        return {"abuse_score":0, "labels":[], "sentiment":"supportive", "stage":4, "darvo_score":0.0, "top_patterns":[]}
-    # model scoring
-    inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True)
-    with torch.no_grad():
-        logits = model(**inputs).logits.squeeze(0)
-    probs = torch.sigmoid(logits).numpy()
-    labels = [lab for lab,p in zip(LABELS,probs) if p > THRESHOLDS[lab]]
-    # abuse score weighted
-    total_w = sum(PATTERN_WEIGHTS.get(l,1.0) for l in LABELS)
-    abuse_score = int(round(sum(probs[i]*PATTERN_WEIGHTS.get(l,1.0) for i,l in enumerate(LABELS))/total_w*100))
+    h=healthy_detector(text)[0]
+    if h['label']=="POSITIVE" and h['score']>0.9:
+        return { "abuse_score":0, "labels":[], "sentiment":"supportive",
+                 "stage":4, "darvo_score":0.0, "top_patterns":[] }
+    # model
+    inp=tokenizer(text,return_tensors='pt',truncation=True,padding=True)
+    with torch.no_grad(): logits=model(**inp).logits.squeeze(0)
+    probs=torch.sigmoid(logits).numpy()
+    labels=[lab for lab,p in zip(LABELS,probs) if p>THRESHOLDS[lab]]
+    # abuse score
+    total_w=sum(PATTERN_WEIGHTS.get(l,1.0) for l in LABELS)
+    abuse_score=int(round(sum(probs[i]*PATTERN_WEIGHTS.get(l,1.0)
+                              for i,l in enumerate(LABELS))/total_w*100))
     # sentiment shift
-    sst = sst_pipeline(text)[0]
-    sentiment = 'supportive' if sst['label']=='POSITIVE' else 'undermining'
-    sent_score = sst['score'] if sentiment=='undermining' else 0.0
+    sst=sst_pipeline(text)[0]
+    sentiment='supportive' if sst['label']=='POSITIVE' else 'undermining'
+    sent_score=sst['score'] if sentiment=='undermining' else 0.0
     # DARVO
     motif_hits, matched = detect_motifs(text)
-    contradiction = detect_contradiction(text)
-    darvo_score = calculate_darvo_score(labels, 0.0, sent_score, matched, contradiction)
+    contradiction=detect_contradiction(text)
+    darvo_score=calculate_darvo_score(labels,0.0,sent_score,matched,contradiction)
     # stage + weapon
-    stage = get_risk_stage(labels, sentiment)
+    stage=get_risk_stage(labels,sentiment)
     if detect_weapon_language(text):
-        abuse_score = min(abuse_score+25,100)
-        stage = max(stage,2)
+        abuse_score=min(abuse_score+25,100)
+        stage=max(stage,2)
     # top patterns
-    top_patterns = sorted(zip(LABELS,probs), key=lambda x: x[1], reverse=True)[:2]
-    return {"abuse_score":abuse_score, "labels":labels, "sentiment":sentiment, "stage":stage, "darvo_score":darvo_score, "top_patterns":top_patterns}
+    top_patterns=sorted(zip(LABELS,probs), key=lambda x:x[1], reverse=True)[:2]
+    return {
+        "abuse_score":abuse_score, "labels":labels, "sentiment":sentiment,
+        "stage":stage, "darvo_score":darvo_score, "top_patterns":top_patterns
+    }
 
-# --- Composite Analysis ---
-def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
-    none_checked = answers_and_none[-1]
-    any_checked = any(answers_and_none[:-1])
-    none_sel = (not any_checked) and none_checked
+# ——— Composite Analysis & UI ———
+def analyze_composite(m1,d1,m2,d2,m3,d3,*answers):
+    none_sel=(answers[-1] and not any(answers[:-1]))
     if none_sel:
-        esc_score = None; risk_level = 'unknown'
+        esc_score=None; risk_level='unknown'
     else:
-        esc_score = sum(w for (_,w),a in zip(ESCALATION_QUESTIONS, answers_and_none[:-1]) if a)
-        risk_level = 'High' if esc_score>=16 else 'Moderate' if esc_score>=8 else 'Low'
-    msgs = [msg1, msg2, msg3]
-    dates = [date1, date2, date3]
-    active = [(m,d) for m,d in zip(msgs, dates) if m.strip()]
+        esc_score=sum(w for (_,w),a in zip(ESCALATION_QUESTIONS,answers[:-1]) if a)
+        risk_level='High' if esc_score>=16 else 'Moderate' if esc_score>=8 else 'Low'
+    msgs=[m1,m2,m3]; dates=[d1,d2,d3]
+    active=[(m,d) for m,d in zip(msgs,dates) if m.strip()]
     if not active:
         return "Please enter at least one message."
-    results = [(analyze_single_message(m), d) for m,d in active]
-    abuse_scores = [r[0]['abuse_score'] for r in results]
-    stage_list = [r[0]['stage'] for r,_ in results]
-    most_common_stage = max(set(stage_list), key=stage_list.count)
-    composite_abuse = int(round(sum(abuse_scores)/len(abuse_scores)))
-    # extract top label from each result dict
-    top_labels = [
-        res['top_patterns'][0][0] 
-        if res['top_patterns'] 
-        else 'None' 
-        for res, _ in results]['top_patterns'][0][0] if r[0]['top_patterns'] else 'None' for r,_ in results
-    ]
-    dates_used = [d or 'Undated' for _,d in results]
-    # common stage
-    stage_list = [r[0]['stage'] for r,_ in results]
-    most_common_stage = max(set(stage_list), key=stage_list.count)
-    composite_abuse = int(round(sum(abuse_scores)/len(abuse_scores)))
+    results=[(analyze_single_message(m),d) for m,d in active]
+    abuse_scores=[r[0]['abuse_score'] for r in results]
+    top_labels=[res['top_patterns'][0][0] if res['top_patterns'] else 'None'
+                for res,_ in results]
+    dates_used=[d or 'Undated' for _,d in results]
+    # common stage & composite abuse
+    stages=[r[0]['stage'] for r,_ in results]
+    most_common_stage=max(set(stages), key=stages.count)
+    composite_abuse=int(round(sum(abuse_scores)/len(abuse_scores)))
     # DARVO summary
-    darvo_vals = [r[0]['darvo_score'] for r,_ in results]
-    avg_darvo = round(sum(darvo_vals)/len(darvo_vals),3)
-    darvo_blurb = ''
-    if avg_darvo>0.25:
-        lvl = 'moderate' if avg_darvo<0.65 else 'high'
-        darvo_blurb = f"\n🎭 DARVO Score: {avg_darvo} ({lvl})"
-    # output text
-    out = f"Abuse Intensity: {composite_abuse}%\n"
+    darvos=[r[0]['darvo_score'] for r,_ in results]
+    avg_darvo=round(sum(darvos)/len(darvos),3)
+    darvo_blurb=f"\n🎭 DARVO Score: {avg_darvo} ({'high' if avg_darvo>=0.65 else 'moderate'})" if avg_darvo>0.25 else ''
+    # build output
+    out=f"Abuse Intensity: {composite_abuse}%\n"
     if esc_score is None:
-        out += "Escalation Potential: Unknown (Checklist not completed)\n"
+        out+="Escalation Potential: Unknown (Checklist not completed)\n"
     else:
-        out += f"Escalation Potential: {risk_level} ({esc_score}/{sum(w for _,w in ESCALATION_QUESTIONS)})\n"
+        out+=f"Escalation Potential: {risk_level} ({esc_score}/{sum(w for _,w in ESCALATION_QUESTIONS)})\n"
     # risk snippet
-    pattern_score = f"{top_labels[0]} – {int(results[0][0]['top_patterns'][0][1]*100)}%" if results[0][0]['top_patterns'] else top_labels[0]
-    out += generate_risk_snippet(composite_abuse, pattern_score, esc_score or 0, most_common_stage)
-    out += darvo_blurb
-    img = generate_abuse_score_chart(dates_used, abuse_scores, top_labels)
+    pattern_score=f"{top_labels[0]} – {int(results[0][0]['top_patterns'][0][1]*100)}%" if results[0][0]['top_patterns'] else top_labels[0]
+    out+=generate_risk_snippet(composite_abuse, pattern_score, esc_score or 0, most_common_stage)
+    out+=darvo_blurb
+    img=generate_abuse_score_chart(dates_used,abuse_scores,top_labels)
     return out, img
 
-# --- Gradio Interface ---
+# ——— Gradio Interface ———
 message_date_pairs = [
     (gr.Textbox(label=f"Message {i+1}"), gr.Textbox(label=f"Date {i+1} (optional)", placeholder="YYYY-MM-DD"))
     for i in range(3)
 ]
-textbox_inputs = [item for pair in message_date_pairs for item in pair]
 quiz_boxes = [gr.Checkbox(label=q) for q,_ in ESCALATION_QUESTIONS]
 none_box = gr.Checkbox(label="None of the above")
+
 iface = gr.Interface(
     fn=analyze_composite,
-    inputs=textbox_inputs + quiz_boxes + [none_box],
+    inputs=[item for pair in message_date_pairs for item in pair] + quiz_boxes + [none_box],
     outputs=[gr.Textbox(label="Results"), gr.Image(label="Risk Stage Timeline", type="pil")],
     title="Tether Abuse Pattern Detector v2",
     allow_flagging="manual"
 )
+
 if __name__ == "__main__":
-    iface.launch()
+    iface.launch()