classifier.py

# classifier.py
import torch
from model_loader import model, tokenizer

def classify_toxic_comment(comment):
    """
    Classify a comment as toxic or non-toxic using the fine-tuned XLM-RoBERTa model.
    Returns the prediction label, confidence, color, toxicity score, and bias score for UI display.
    """
    if not comment.strip():
        return "Error: Please enter a comment.", None, None, None, None

    # Tokenize the input comment
    inputs = tokenizer(comment, return_tensors="pt", truncation=True, padding=True, max_length=512)

    # Run inference
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits

    # Get the predicted class (0 = non-toxic, 1 = toxic)
    predicted_class = torch.argmax(logits, dim=1).item()
    label = "Toxic" if predicted_class == 1 else "Non-Toxic"
    confidence = torch.softmax(logits, dim=1)[0][predicted_class].item()
    label_color = "red" if label == "Toxic" else "green"

    # Simulate Toxicity Score (in a real scenario, use a model like Detoxify)
    # For now, we'll approximate it based on the confidence of the toxic class
    toxicity_score = torch.softmax(logits, dim=1)[0][1].item()  # Probability of toxic class
    toxicity_score = round(toxicity_score, 2)

    # Simulate Bias Score (in a real scenario, use a bias detection model like WEAT)
    # For now, we'll use a placeholder value (since the example comment is non-toxic)
    bias_score = 0.01 if label == "Non-Toxic" else 0.15  # Placeholder logic
    bias_score = round(bias_score, 2)

    return f"Prediction: {label}", confidence, label_color, toxicity_score, bias_score