app.py

import gradio as gr
import torch
from transformers import RobertaForSequenceClassification, RobertaTokenizer
import numpy as np
from transformers import pipeline

# Load sentiment analysis model
sentiment_analyzer = pipeline("sentiment-analysis")

# Load model and tokenizer
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": 0.30,
    "control": 0.43, "guilt_tripping": 0.19, "apology_baiting": 0.45,
    "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": 0.70,
    "suicidal_threat": 0.45, "physical_threat": 0.20, "extreme_control": 0.36
}

PATTERN_LABELS = LABELS[:15]
DANGER_LABELS = LABELS[15:]

EXPLANATIONS = {
    "gaslighting": "Gaslighting involves making someone question their own reality or perceptions.",
    "blame_shifting": "Blame-shifting is when one person redirects responsibility onto someone else.",
    "projection": "Projection accuses the victim of behaviors the abuser exhibits themselves.",
    "dismissiveness": "Dismissiveness belittles or ignores another person’s thoughts or feelings.",
    "mockery": "Mockery involves ridicule or sarcasm meant to humiliate.",
    "recovery_phase": "Recovery phase invalidates someone’s healing process or needs.",
    "insults": "Insults are derogatory remarks meant to degrade or attack.",
    "apology_baiting": "Apology-baiting manipulates someone into apologizing for the abuser’s actions.",
    "deflection": "Deflection shifts responsibility or changes the subject to avoid blame.",
    "control": "Control includes behavior that limits another’s autonomy or freedom.",
    "extreme_control": "Extreme control is highly manipulative dominance over another’s choices.",
    "physical_threat": "Physical threats suggest or state a risk of bodily harm.",
    "suicidal_threat": "Suicidal threats use self-harm as a way to manipulate others.",
    "guilt_tripping": "Guilt-tripping makes someone feel guilty for things they didn’t cause.",
    "manipulation": "Manipulation influences someone’s behavior through deceptive emotional tactics.",
    "non_abusive": "Non-abusive language is respectful, supportive, and healthy."
}

def calculate_abuse_level(scores, thresholds):
    triggered_scores = [score for label, score in zip(LABELS, scores) if score > thresholds[label]]
    if not triggered_scores:
        return 0.0
    return round(np.mean(triggered_scores) * 100, 2)

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"
    else:
        return "Very Low / Likely Safe"

def analyze_messages(input_text, context_flags):
    input_text = input_text.strip()
    if not input_text:
        return "Please enter a message for analysis."

    # Sentiment
    sentiment = sentiment_analyzer(input_text)[0]
    sentiment_label = sentiment['label']
    sentiment_score = sentiment['score']

    # Threshold adjustment for negative tone
    adjusted_thresholds = THRESHOLDS.copy()
    if sentiment_label == "NEGATIVE":
        adjusted_thresholds = {key: val * 0.8 for key, val in THRESHOLDS.items()}

    # Tokenization and prediction
    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()

    # Pattern and danger flags from model
    pattern_count = sum(score > adjusted_thresholds[label] for label, score in zip(PATTERN_LABELS, scores[:15]))
    danger_flag_count = sum(score > adjusted_thresholds[label] for label, score in zip(DANGER_LABELS, scores[15:]))

    # Add contextual danger from checkboxes
    if context_flags and len(context_flags) >= 2:
        danger_flag_count += 1

    # Non-abusive override
    non_abusive_score = scores[LABELS.index('non_abusive')]
    if non_abusive_score > adjusted_thresholds['non_abusive']:
        return "This message is classified as non-abusive."

    # Abuse level
    abuse_level = calculate_abuse_level(scores, THRESHOLDS)
    abuse_description = interpret_abuse_level(abuse_level)

    # Resources
    if danger_flag_count >= 2:
        resources = "⚠️ Your responses indicate elevated danger. Please consider seeking support immediately through a domestic violence hotline or trusted professional."
    else:
        resources = "For more information on abuse patterns, consider reaching out to support groups or professional counselors."

    # Top patterns with definitions
    scored_patterns = [(label, score) for label, score in zip(PATTERN_LABELS, scores[:15])]
    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
    ])

    result = (
        f"Abuse Risk Score: {abuse_level}% – {abuse_description}\n\n"
        f"Most Likely Patterns:\n{top_pattern_explanations}\n\n"
        f"⚠️ Critical Danger Flags Detected: {danger_flag_count} of 3\n"
        "The Danger Assessment is a validated tool that helps identify serious risk in intimate partner violence. "
        "It flags communication patterns associated with increased risk of severe harm.\n\n"
        f"Resources: {resources}\n\n"
        f"Sentiment: {sentiment_label} (Confidence: {sentiment_score*100:.2f}%)"
    )

    return result

# Launch interface
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",
)

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