app.py

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
sentiment_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-sentiment")
sentiment_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-sentiment")

# Load abuse pattern model
model_name = "SamanthaStorm/abuse-pattern-detector-v2"
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",
    "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 = {
    "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."
}

def custom_sentiment(text):
    inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    with torch.no_grad():
        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}

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  # shouldn't contribute to abuse score
}

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)
    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"}):
        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"
    return "Very Low / Likely Safe"

def analyze_messages(input_text, risk_flags):
    input_text = input_text.strip()
    if not input_text:
        return "Please enter a message for analysis."
    
    # Normalize the text (example: lower case)
    normalized_text = input_text.strip().lower()

    motif_flags, matched_phrases = detect_motifs(input_text)
    risk_flags = list(set(risk_flags + motif_flags)) if risk_flags else motif_flags

    sentiment = custom_sentiment(input_text)
    sentiment_label = sentiment['label']
    sentiment_score = sentiment['score']

    adjusted_thresholds = {k: v * 0.8 for k, v in THRESHOLDS.items()} if sentiment_label == "undermining" else THRESHOLDS.copy()

    inputs = tokenizer(input_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 > adjusted_thresholds[label]]
    phrase_labels = [label for label, _ in matched_phrases]
    pattern_labels_used = list(set(threshold_labels + phrase_labels))
  
    contextual_flags = risk_flags if risk_flags else []
    # Note: If there are two or more contextual flags, you might wish to adjust a danger counter
    # danger_flag_count += 1  <-- Ensure that danger_flag_count is defined before incrementing.
    abuse_level = calculate_abuse_level(scores, adjusted_thresholds, motif_hits=[label for label, _ in matched_phrases])
    abuse_description = interpret_abuse_level(abuse_level)
    
    # Escalate risk if user checks a critical context box
    if contextual_flags and abuse_level < 15:
        abuse_level = 15  # bump to at least Mild Concern

    abuse_type, abuser_profile, advice = determine_abuse_type(pattern_labels_used)

    danger_flag_count = sum(score > adjusted_thresholds[label] for label, score in zip(DANGER_LABELS, scores[15:18]))

    critical_flags = ["They've threatened harm", "They monitor/follow me", "I feel unsafe when alone with them"]
    high_risk_context = any(flag in contextual_flags for flag in critical_flags)

    non_abusive_score = scores[LABELS.index('non_abusive')]
    non_abusive_confident = non_abusive_score > adjusted_thresholds['non_abusive']

    if non_abusive_confident and danger_flag_count == 0 and not matched_phrases:
        return "This message is classified as non-abusive."
    # Supportive override logic
    if (
        sentiment_label == "supportive"
            and sentiment_score > 0.95
            and non_abusive_confident
            and danger_flag_count == 0
            and not matched_phrases
    ):
        return "This message is classified as non-abusive. It appears emotionally supportive and safe."
    
    scored_patterns = [
        (label, score) for label, score in zip(PATTERN_LABELS, scores[:15]) if label != "non_abusive"
    ]

    override_labels = {"physical_threat", "suicidal_threat", "extreme_control"}
    override_matches = [label for label, _ in matched_phrases if label in override_labels]

    if override_matches:
        top_patterns = [(label, 1.0) for label in override_matches]
    else:
        top_patterns = sorted(scored_patterns, key=lambda x: x[1], reverse=True)[:2]

    top_pattern_explanations = "\n".join([
        f"• {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label, 'No explanation available.')}"
        for label, _ in top_patterns
    ])

    resources = "Immediate assistance recommended. Please seek professional help or contact emergency services." if danger_flag_count >= 2 else "For more information on abuse patterns, consider reaching out to support groups or professional counselors."

    result = f"Abuse Risk Score: {abuse_level}% – {abuse_description}\n\n"
    if abuse_level >= 15:
        result += f"Most Likely Patterns:\n{top_pattern_explanations}\n\n"
        result += f"⚠️ Critical Danger Flags Detected: {danger_flag_count} of 3\n"
        result += f"Resources: {resources}\n"
        result += f"🧠 Sentiment: {sentiment_label.title()} (Confidence: {sentiment_score*100:.2f}%)\n"

    if contextual_flags:
        result += "\n\n⚠️ You indicated the following:\n" + "\n".join([f"• {flag.replace('_', ' ').title()}" for flag in contextual_flags])

    if high_risk_context:
        result += "\n\n🚨 These responses suggest a high-risk situation. Consider seeking immediate help or safety planning resources."

    if matched_phrases:
        result += "\n\n🚨 Detected High-Risk Phrases:\n"
        for label, phrase in matched_phrases:
            phrase_clean = phrase.replace('"', "'").strip()
            result += f"• {label.replace('_', ' ').title()}: “{phrase_clean}”\n"

    if abuse_type:
        result += f"\n\n🧠 Likely Abuse Type: {abuse_type}"
        result += f"\n🧠 Abuser Profile: {abuser_profile}"
        result += f"\n📘 Safety Tip: {advice}"

    return result

# Updated Interface: Added flagging functionality to allow users to flag mispredictions.
iface = gr.Interface(
    fn=analyze_messages,
    inputs=[
        gr.Textbox(lines=10, placeholder="Enter message here..."),
        gr.CheckboxGroup(label="Do any of these apply to your situation?", 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="Analysis Result")],
    title="Abuse Pattern Detector",
    live=True,
    allow_flagging="manual"  # This enables the manual flagging button for user feedback.
)

if __name__ == "__main__":
    iface.queue().launch()