app.py

import os
import logging
from typing import Dict, Any

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


def predict(text: str) -> Dict[str, Any]:
    """Classify text for PII detection."""
    if not text or text.strip() == "":
        return {"No input provided": 0.0}
    logging.info(f"User input: {text}")

    try:
        # Tokenize input
        inputs = tokenizer(
            text,
            return_tensors="pt",
            padding="max_length",
            max_length=512,
            truncation=True
        )

        # Run inference
        with torch.no_grad():
            outputs = model(**inputs)
            logits = outputs.logits
            probabilities = torch.sigmoid(logits)
            probs = probabilities.squeeze().tolist()

        # Create results dictionary
        results = {
            "Asking for PII": float(probs[0]),
            "Giving PII": float(probs[1])
        }

        return results

    except Exception as e:
        return {"Error": str(e)}


# Example test cases
examples = [
    ["what's your blue app id?"],
    ["I live at 901 Roosevelt St, Redwood City"],
    ["what's you ph0ne rebmun?"],
    ["yellow gh>>ost app id? let's chat there"],
    ["let's z0000m?"],
    ["Let’s meet at the Starbuck close to Stanford"],
]


if __name__ == "__main__":
    logging.basicConfig(
        level=logging.INFO,
        format='%(asctime)s - %(levelname)s - %(message)s',
        handlers=[logging.StreamHandler()]
    )
    
    # Model configuration
    model_id = "Roblox/roblox-pii-classifier"

    # Get HF token from Hugging Face Space secrets
    # In Spaces, set HF_TOKEN in Settings > Repository secrets
    HF_TOKEN = os.getenv("HF_TOKEN")

    # Load model and tokenizer
    print(f"Loading model: {model_id}")
    try:
        # Use token if available (required for private models)
        if HF_TOKEN:
            print("Using HF_TOKEN from environment/secrets")
            model = AutoModelForSequenceClassification.from_pretrained(model_id, token=HF_TOKEN)
            tokenizer = AutoTokenizer.from_pretrained(model_id, token=HF_TOKEN)
        else:
            print("No HF_TOKEN found, attempting without authentication...")
            model = AutoModelForSequenceClassification.from_pretrained(model_id)
            tokenizer = AutoTokenizer.from_pretrained(model_id)

        model.eval()
        print("Model loaded successfully!")
    except Exception as e:
        print(f"Failed to load model: {e}")
        if not HF_TOKEN:
            print("\n⚠️  For private models, you need to set HF_TOKEN as a Space secret:")
            print("   1. Go to your Space Settings")
            print("   2. Add a new secret named 'HF_TOKEN'")
            print("   3. Set your Hugging Face token as the value")
        exit(1)

    # Create Gradio interface
    demo = gr.Interface(
        fn=predict,
        inputs=gr.Textbox(
            lines=3,
            placeholder="Enter text to analyze for PII content...",
            label="Input Text"
        ),
        outputs=gr.Label(
            num_top_classes=2,
            label="Classification Results"
        ),
        title="PII Detection Demo",
        description="This model detects whether text is asking for or giving personal information (PII).",
        examples=examples,
        flagging_mode="never",
    )

    demo.launch()