Update app.py

app.py

@@ -9,8 +9,40 @@ import io
 from PIL import Image
 from datetime import datetime
 
-# ——— DARVO & Risk Utilities ———
-DARVO_PATTERNS = {"blame shifting", "projection", "dismissiveness", "guilt tripping", "contradictory statements"}
+# ——— Constants & Utilities ———
+LABELS = [
+    "blame shifting", "contradictory statements", "control", "dismissiveness",
+    "gaslighting", "guilt tripping", "insults", "obscure language",
+    "projection", "recovery phase", "threat"
+]
+
+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
+}
+
+PATTERN_WEIGHTS = {
+    "gaslighting": 1.3,
+    "control": 1.2,
+    "dismissiveness": 0.8,
+    "blame shifting": 0.8,
+    "contradictory statements": 0.75,
+    "threat": 1.5
+}
+
+DARVO_PATTERNS = set([
+    "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.",
@@ -27,33 +59,30 @@ DARVO_MOTIFS = [
     "You’re the one who’s always complaining.", "You’re the one who’s always starting arguments.",
     "You’re the one who’s always making things worse.", "You’re the one who’s always making me feel bad.",
     "You’re the one who’s always making me look like the bad guy.",
-    "You’re the one who’s always making me feel like a failure.", "You’re the one who’s always making me feel like I’m not good enough.",
-    "I can’t believe you’re doing this to me.", "You’re hurting me.", "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 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 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."
+    "You’re the one who’s always making me feel like a failure.",
+    "You’re the one who’s always making me feel like I’m not good enough.",
+    "I can’t believe you’re doing this to me.", "You’re hurting me.",
+    "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 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."
 ]
-PATTERN_WEIGHTS = {
-    "gaslighting": 1.3,
-    "control": 1.2,
-    "dismissiveness": 0.8,
-    "blame shifting": 0.8,
-    "contradictory statements": 0.75,
-    "threat": 1.5,
-}
+
 RISK_STAGE_LABELS = {
     1: "🌀 Risk Stage: Tension-Building\nThis message reflects rising emotional pressure or subtle control attempts.",
     2: "🔥 Risk Stage: Escalation\nThis message includes direct or aggressive patterns, suggesting active harm.",
     3: "🌧️ Risk Stage: Reconciliation\nThis message reflects a reset attempt—apologies or emotional repair without accountability.",
     4: "🌸 Risk Stage: Calm / Honeymoon\nThis message appears supportive but may follow prior harm, minimizing it."
 }
+
 ESCALATION_QUESTIONS = [
     ("Partner has access to firearms or weapons", 4),
     ("Partner threatened to kill you", 3),
@@ -71,7 +100,10 @@ def detect_contradiction(message):
     patterns = [
         (r"\b(i love you).{0,15}(i hate you|you ruin everything)", re.IGNORECASE),
         (r"\b(i’m sorry).{0,15}(but you|if you hadn’t)", re.IGNORECASE),
-        # ... other patterns ...
+        (r"\b(i’m trying).{0,15}(you never|why do you)", re.IGNORECASE),
+        (r"\b(do what you want).{0,15}(you’ll regret it|i always give everything)", re.IGNORECASE),
+        (r"\b(i don’t care).{0,15}(you never think of me)", re.IGNORECASE),
+        (r"\b(i guess i’m just).{0,15}(the bad guy|worthless|never enough)", re.IGNORECASE)
     ]
     return any(re.search(pat, message, flags) for pat, flags in patterns)
 
@@ -90,116 +122,151 @@ def generate_risk_snippet(abuse_score, top_label, escalation_score, stage):
     why = {
         "control": "This message may reflect efforts to restrict someone’s autonomy.",
         "gaslighting": "This message could be manipulating perception.",
-        # ... other explanations ...
+        "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.")
     if abuse_score>=85 or escalation_score>=16:
-        lvl="high"
+        lvl = "high"
     elif abuse_score>=60 or escalation_score>=8:
-        lvl="moderate"
+        lvl = "moderate"
     else:
-        lvl="low"
-    return f"\n\n🛑 Risk Level: {lvl.capitalize()}\nThis message shows **{label}**.\n💡 Why: {why}\n"
+        lvl = "low"
+    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"]
-    t=text.lower()
+    kws = ["knife","gun","bomb","kill you","shoot","explode"]
+    t = text.lower()
     return any(k in t for k in kws)
 
 def get_risk_stage(patterns, sentiment):
-    if "threat" in patterns or "insults" in patterns: return 2
-    if "control" in patterns or "guilt tripping" in patterns: return 1
-    if "recovery phase" in patterns: return 3
-    if sentiment=="supportive" and any(p in patterns for p in ["projection","dismissiveness"]): return 4
+    if "threat" in patterns or "insults" in patterns:
+        return 2
+    if "control" in patterns or "guilt tripping" in patterns:
+        return 1
+    if "recovery phase" in patterns:
+        return 3
+    if sentiment == "supportive" and any(p in patterns for p in ["projection","dismissiveness"]):
+        return 4
     return 1
 
-# --- Timeline Visualization ---
+# --- Visualization ---
 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 = list(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",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)
+        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)
     return Image.open(buf)
 
-# --- Load and initialize models ---
-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")
-
-LABELS=[
-    "blame shifting","contradictory statements","control","dismissiveness",
-    "gaslighting","guilt tripping","insults","obscure language",
-    "projection","recovery phase","threat"
-]
-THRESHOLDS={l:THRESHOLDS.get(l,0.3) for l in LABELS}
+# --- 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):
-    if healthy_detector(text)[0]['label']=="POSITIVE" and healthy_detector(text)[0]['score']>0.9:
-        return {"abuse_score":0,"labels":[],"sentiment":"supportive","stage":4,"darvo_score":0.0,"top_patterns":[]}
-    inputs=tokenizer(text,return_tensors='pt',padding=True,truncation=True)
-    with torch.no_grad(): outputs=model(**inputs).logits.squeeze(0)
-    probs=torch.sigmoid(outputs).numpy()
-    labels=[lab for lab,p in zip(LABELS,probs) if p>THRESHOLDS[lab]]
-    # weighted 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
-    sst=sst_pipeline(text)[0]
-    sentiment='supportive' if sst['label']=='POSITIVE' else 'undermining'
-    sent_score=sst['score'] if sentiment=='undermining' else 0.0
+    # 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))
+    # 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
     # DARVO
-    motifs,matched=detect_motifs(text)
-    contradiction=detect_contradiction(text)
-    darvo=calculate_darvo_score(labels,0.0,sent_score,matched,contradiction)
+    motif_hits, matched = detect_motifs(text)
+    contradiction = detect_contradiction(text)
+    darvo_score = calculate_darvo_score(labels, 0.0, sent_score, matched, contradiction)
     # stage + weapon
-    stage=get_risk_stage(labels,sentiment)
-    if detect_weapon_language(text): abuse_score=min(abuse_score+25,100); stage=max(stage,2)
+    stage = get_risk_stage(labels, sentiment)
+    if detect_weapon_language(text):
+        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,"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(m1,d1,m2,d2,m3,d3,*answers):
-    none_sel=answers[-1] and not any(answers[:-1])
-    if none_sel: esc=None; risk='unknown'
+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
+    if none_sel:
+        esc_score = None; risk_level = 'unknown'
     else:
-        esc=sum(w for (_,w),a in zip(ESCALATION_QUESTIONS,answers[:-1]) if a)
-        risk='High' if esc>=16 else 'Moderate' if esc>=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]
-    top_lbls=[r[0]['top_patterns'][0][0] if r[0]['top_patterns'] else 'None' for r in results]
-    dates_used=[d or 'Undated' for (_,d) 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)))
-    out=f"Abuse Intensity: {composite_abuse}%\n"
-    if esc is None: out+="Escalation Potential: Unknown (Checklist not completed)\n"
-    else: out+=f"Escalation Potential: {risk} ({esc}/{sum(w for _,w in ESCALATION_QUESTIONS)})\n"
+        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()]
+    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]
+    top_labels = [r[0]['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)))
     # DARVO summary
-    darvos=[r[0]['darvo_score'] for r,_ in results]
-    avg_d=sum(darvos)/len(darvos)
-    if avg_d>0.25: lvl='moderate' if avg_d<0.65 else 'high'; out+=f"\n🎭 DARVO Score: {round(avg_d,3)} ({lvl})\n"
-    out+=generate_risk_snippet(composite_abuse,f"{top_lbls[0]} – {int(top_patterns[0][1]*100)}%",esc or 0,most_common_stage)
-    img=generate_abuse_score_chart(dates_used,abuse_scores,top_lbls)
-    return out,img
-
-# --- UI ---
-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)]
-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=[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()
+    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"
+    if esc_score is None:
+        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"
+    # 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)
+    return out, img
+
+# --- 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],
+    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()
+