app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import numpy as np

# Load bias detection model
model_name = "d4data/bias-detection-model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Bias categories
bias_labels = ["Non-biased", "Biased"]

def analyze_bias(text):
    """
    Analyze text for potential bias using a fine-tuned transformer model.
    Returns bias classification, confidence score, and detailed analysis.
    """
    if not text or len(text.strip()) == 0:
        return "Please enter some text to analyze.", "", ""
    
    # Tokenize and predict
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512, padding=True)
    
    with torch.no_grad():
        outputs = model(**inputs)
        predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    
    # Get prediction results
    confidence_scores = predictions[0].tolist()
    predicted_class = torch.argmax(predictions, dim=-1).item()
    
    # Generate detailed analysis
    bias_score = confidence_scores[1] * 100  # Biased probability
    neutral_score = confidence_scores[0] * 100  # Non-biased probability
    
    # Classification result
    if predicted_class == 1:
        result = f"⚠️ **BIAS DETECTED** (Confidence: {bias_score:.1f}%)"
        interpretation = "This text shows signs of potential bias. Consider reviewing for:"
        recommendations = [
            "• Language that may unfairly represent certain groups",
            "• Stereotypical assumptions or generalizations",
            "• Exclusionary or prejudiced framing",
            "• Consider rephrasing with more neutral, inclusive language"
        ]
    else:
        result = f"✅ **NEUTRAL TEXT** (Confidence: {neutral_score:.1f}%)"
        interpretation = "This text appears relatively neutral and unbiased."
        recommendations = [
            "• Continue using inclusive, balanced language",
            "• Maintain awareness of potential implicit biases",
            "• Consider diverse perspectives in your writing"
        ]
    
    # Detailed breakdown
    breakdown = f"""### Bias Analysis Scores:
- **Biased**: {bias_score:.2f}%
- **Neutral**: {neutral_score:.2f}%

### {interpretation}
{''.join([f'{rec}\n' for rec in recommendations])}

### About This Tool:
This AI bias detector uses advanced natural language processing to identify potential biases in text. 
It's designed for researchers, journalists, educators, and content creators who want to ensure 
fair and inclusive communication.
    """
    
    return result, breakdown, f"Bias Score: {bias_score:.1f}% | Neutral Score: {neutral_score:.1f}%"

# Create Gradio interface
with gr.Blocks(title="AI Bias Detector", theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        """
        # 🎯 AI Bias Detector
        ### Award-Winning Tool for Detecting Bias in Text and AI Outputs
        
        A powerful AI tool designed for researchers, journalists, and educators to detect potential bias 
        in text. Features cutting-edge analysis, user-friendly UI, and insightful visualizations.
        
        **How to use:** Enter or paste text below, then click "Analyze Bias" to receive a comprehensive analysis.
        """
    )
    
    with gr.Row():
        with gr.Column(scale=1):
            text_input = gr.Textbox(
                label="Enter Text to Analyze",
                placeholder="Paste or type text here to check for potential bias...",
                lines=10,
                max_lines=20
            )
            
            analyze_btn = gr.Button("🔍 Analyze Bias", variant="primary", size="lg")
            
            gr.Examples(
                examples=[
                    ["The scientist made a groundbreaking discovery in quantum physics."],
                    ["Women are naturally better at multitasking than men."],
                    ["The CEO announced new initiatives to promote workplace diversity and inclusion."],
                    ["All teenagers are lazy and obsessed with social media."],
                    ["The research team, comprising experts from diverse backgrounds, published their findings."]
                ],
                inputs=text_input,
                label="Example Texts (Click to Try)"
            )
        
        with gr.Column(scale=1):
            result_output = gr.Markdown(label="Analysis Result")
            score_output = gr.Textbox(label="Summary", interactive=False)
            detailed_output = gr.Markdown(label="Detailed Analysis")
    
    analyze_btn.click(
        fn=analyze_bias,
        inputs=text_input,
        outputs=[result_output, detailed_output, score_output]
    )
    
    gr.Markdown(
        """
        ---
        ### 📊 About This Tool
        
        This bias detector uses state-of-the-art transformer models fine-tuned on bias detection datasets. 
        It analyzes linguistic patterns, word choices, and contextual meanings to identify potential biases.
        
        **Note:** This tool provides suggestions and should be used as part of a comprehensive editorial review process. 
        Human judgment remains essential for nuanced content evaluation.
        
        **Developed by:** airifoundation | **Model:** d4data/bias-detection-model
        """
    )

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