app.py

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

# Load model and tokenizer with trust_remote_code in case it's needed
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)

# Define labels (17 total)
LABELS = [
    "gaslighting", "mockery", "dismissiveness", "control",
    "guilt_tripping", "apology_baiting", "blame_shifting", "projection",
    "contradictory_statements", "manipulation", "deflection", "insults",
    "obscure_formal", "recovery_phase", "suicidal_threat", "physical_threat",
    "extreme_control"
]

# Custom thresholds for each label (make sure these match your original settings)
THRESHOLDS = {
    "gaslighting": 0.25,
    "mockery": 0.15,
    "dismissiveness": 0.30,  # original value, not 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,
    "suicidal_threat": 0.45,
    "physical_threat": 0.31,
    "extreme_control": 0.36,
    "non_abusive": 0.40
}

# Define label groups using slicing (first 14: abuse patterns, last 3: danger cues)
PATTERN_LABELS = LABELS[:14]
DANGER_LABELS = LABELS[14:17]

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):
    input_text = input_text.strip()
    if not input_text:
        return "Please enter a message for analysis.", None

    # Tokenize input and generate model predictions
    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()

    # Count the number of triggered abuse pattern and danger flags based on thresholds
    pattern_count = sum(score > THRESHOLDS[label] for label, score in zip(PATTERN_LABELS, scores[:14]))
    danger_flag_count = sum(score > THRESHOLDS[label] for label, score in zip(DANGER_LABELS, scores[14:17]))

    # Build formatted raw score display
    score_lines = [
        f"{label:25}: {score:.3f}" for label, score in zip(PATTERN_LABELS + DANGER_LABELS, scores)
    ]
    raw_score_output = "\n".join(score_lines)

    # Calculate overall abuse level and interpret it
    abuse_level = calculate_abuse_level(scores, THRESHOLDS)
    abuse_description = interpret_abuse_level(abuse_level)

    # Resource logic based on the number of danger cues
    if danger_flag_count >= 2:
        resources = "Immediate assistance recommended. Please seek professional help or contact emergency services."
    else:
        resources = "For more information on abuse patterns, consider reaching out to support groups or professional counselors."

    # Get top 2 highest scoring abuse patterns (excluding 'non_abusive')
    scored_patterns = [(label, score) for label, score in zip(PATTERN_LABELS, scores[:14])]
    top_patterns = sorted(scored_patterns, key=lambda x: x[1], reverse=True)[:2]
    top_patterns_str = "\n".join([f"• {label.replace('_', ' ').title()}" for label, _ in top_patterns])

    # Format final result
    result = (
        f"Abuse Risk Score: {abuse_level}% – {abuse_description}\n"
        "This message contains signs of emotionally harmful communication that may indicate abusive patterns.\n\n"
        f"Most Likely Patterns:\n{top_patterns_str}\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. "
        "For more info, consider reaching out to support groups or professionals.\n\n"
        f"Resources: {resources}"
    )

    # Return both a text summary and a JSON-like dict of scores per label
    return result

# Updated Gradio Interface using new component syntax
iface = gr.Interface(
    fn=analyze_messages,
    inputs=gr.Textbox(lines=10, placeholder="Enter message here..."),
    outputs=[
        gr.Textbox(label="Analysis Result"),
    ],
    title="Abuse Pattern Detector"
)

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