app.py

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

# Load more accurate detection model
MODEL = "desklib/ai-text-detector-v1.01"
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForSequenceClassification.from_pretrained(MODEL)

def get_color(ai_score):
    red = int(ai_score * 255)
    green = int((1 - ai_score) * 255)
    return f"rgb({red},{green},0)"

def detect_ai(text):
    sentences = re.split(r'(?<=[.!?]) +', text)
    results = []

    for sent in sentences:
        if not sent.strip():
            continue
        inputs = tokenizer(sent, return_tensors="pt", truncation=True, max_length=512)
        with torch.no_grad():
            outputs = model(**inputs)
            probs = torch.softmax(outputs.logits, dim=1)
        ai_score = float(probs[0][1])
        results.append({"sentence": sent, "ai_score": ai_score})

    highlighted = ""
    for r in results:
        color = get_color(r['ai_score'])
        highlighted += f"<span style='background-color:{color}; padding:2px'>{r['sentence']} </span>"

    if results:
        avg_ai = sum(r['ai_score'] for r in results) / len(results)
        total_percent = round(avg_ai * 100, 2)
        highlighted += f"<p><b>Total AI Probability: {total_percent}%</b></p>"
    else:
        total_percent = 0.0

    return highlighted, {"sentences": results, "total_ai_percent": total_percent}

with gr.Blocks() as demo:
    gr.Markdown("##  AI Detector (upgraded to DeBERTa-v3 large)")
    gr.Markdown("Paste text: green = human-like, yellow = mixed, red = AI-like.")
    input_text = gr.Textbox(lines=8, placeholder="Enter text here…")
    output_html = gr.HTML()
    output_json = gr.JSON()
    run_btn = gr.Button("Detect AI")
    run_btn.click(detect_ai, inputs=input_text, outputs=[output_html, output_json])

demo.launch()