app.py

"""
Privacy Classifier Demo - Classifies prompts as KEEP_LOCAL vs ALLOW_CLOUD
"""

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

# Model configuration
MODEL_ID = "jonmabe/privacy-classifier-electra"
# Model labels: 0=safe (ALLOW_CLOUD), 1=sensitive (KEEP_LOCAL)
LABELS = ["ALLOW_CLOUD", "KEEP_LOCAL"]  # index 0=safe, index 1=sensitive

# Load model and tokenizer
print("Loading model...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
model.eval()

# Move to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
print(f"Model loaded on {device}")


def classify_prompt(text: str) -> tuple[str, dict]:
    """
    Classify a prompt as KEEP_LOCAL or ALLOW_CLOUD.
    
    Returns:
        - Classification label with confidence
        - Dictionary of class probabilities for the label component
    """
    if not text.strip():
        return "Please enter a prompt to classify.", {}
    
    # Tokenize
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=512,
        padding=True
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    # Inference
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.softmax(logits, dim=-1)[0]
    
    # Get prediction
    pred_idx = probs.argmax().item()
    pred_label = LABELS[pred_idx]
    confidence = probs[pred_idx].item()
    
    # Create probability dict for label component
    prob_dict = {label: float(probs[i]) for i, label in enumerate(LABELS)}
    
    return f"{pred_label} ({confidence:.1%} confidence)", prob_dict


def get_color_for_label(label: str) -> str:
    """Return color based on classification."""
    if "KEEP_LOCAL" in label:
        return "red"
    elif "ALLOW_CLOUD" in label:
        return "green"
    return "gray"


# Example prompts
EXAMPLES = [
    ["What is the capital of France?"],
    ["My social security number is 123-45-6789, can you help me file taxes?"],
    ["Write me a poem about the ocean."],
    ["Here's my password: hunter2, please remember it."],
    ["Explain how photosynthesis works."],
    ["My credit card number is 4111-1111-1111-1111, check if it's valid."],
    ["What are some good restaurants in Seattle?"],
    ["My medical records show I have diabetes. What should I eat?"],
    ["Translate 'hello world' to Spanish."],
    ["My home address is 123 Main St, Anytown USA. Send me a pizza."],
    ["How do I sort a list in Python?"],
    ["My employee ID is E12345 and my salary is $85,000."],
]


# Custom CSS for styling
css = """
.keep-local {
    background: linear-gradient(135deg, #ff6b6b 0%, #ee5a5a 100%) !important;
    color: white !important;
    font-weight: bold !important;
}
.allow-cloud {
    background: linear-gradient(135deg, #51cf66 0%, #40c057 100%) !important;
    color: white !important;
    font-weight: bold !important;
}
.result-box {
    font-size: 1.2em;
    padding: 20px;
    border-radius: 10px;
    text-align: center;
}
"""

# Create Gradio interface
with gr.Blocks(css=css, title="Privacy Classifier") as demo:
    gr.Markdown("""
    # 🔒 Privacy Classifier
    
    Classify prompts to determine if they contain sensitive information that should stay local
    or if they're safe to send to cloud LLM services.
    
    - **🔴 KEEP_LOCAL**: Contains PII, sensitive data, or private information
    - **🟢 ALLOW_CLOUD**: Safe to process with cloud-based AI services
    
    This model helps route requests in privacy-aware AI systems.
    """)
    
    with gr.Row():
        with gr.Column(scale=2):
            input_text = gr.Textbox(
                label="Enter your prompt",
                placeholder="Type a prompt to classify...",
                lines=3,
            )
            classify_btn = gr.Button("🔍 Classify", variant="primary", size="lg")
        
        with gr.Column(scale=1):
            result_label = gr.Textbox(
                label="Classification Result",
                interactive=False,
                lines=2,
            )
            confidence_chart = gr.Label(
                label="Confidence Scores",
                num_top_classes=2,
            )
    
    gr.Markdown("### 📝 Example Prompts")
    gr.Examples(
        examples=EXAMPLES,
        inputs=input_text,
        outputs=[result_label, confidence_chart],
        fn=classify_prompt,
        cache_examples=False,
    )
    
    # Event handlers
    classify_btn.click(
        fn=classify_prompt,
        inputs=input_text,
        outputs=[result_label, confidence_chart],
    )
    
    input_text.submit(
        fn=classify_prompt,
        inputs=input_text,
        outputs=[result_label, confidence_chart],
    )
    
    gr.Markdown("""
    ---
    ### About This Model
    
    **Model**: [jonmabe/privacy-classifier-electra](https://huggingface.co/jonmabe/privacy-classifier-electra)
    
    This is an ELECTRA-based classifier fine-tuned to detect sensitive information in prompts.
    Use cases include:
    - Privacy-aware prompt routing
    - Data loss prevention for LLM applications
    - Compliance with data protection regulations
    
    ⚠️ **Disclaimer**: This model is for demonstration purposes. Always verify classifications
    for production use cases involving sensitive data.
    """)


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