Update app.py

app.py

@@ -1,45 +1,61 @@
-from datasets import load_dataset
-from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
+!pip install gradio transformers torch
+
 import gradio as gr
+from transformers import AutoTokenizer, AutoModelForSequenceClassification, GPT2LMHeadModel
+import torch
+import math
+
+# Load a free AI detector model (RoBERTa)
+detector_name = "Hello-SimpleAI/AI-Text-Detector-RoBERTa"
+detector_tokenizer = AutoTokenizer.from_pretrained(detector_name)
+detector_model = AutoModelForSequenceClassification.from_pretrained(detector_name)
+
+# Load GPT-2 for perplexity scoring
+gpt2_name = "gpt2"
+gpt2_tokenizer = AutoTokenizer.from_pretrained(gpt2_name)
+gpt2_model = GPT2LMHeadModel.from_pretrained(gpt2_name)
+
+def compute_perplexity(text):
+    enc = gpt2_tokenizer(text, return_tensors="pt")
+    input_ids = enc.input_ids
+    with torch.no_grad():
+        loss = gpt2_model(input_ids, labels=input_ids).loss
+    return math.exp(loss.item())
+
+def analyze_text(user_text):
+    # Model 1: RoBERTa detector
+    inputs = detector_tokenizer(user_text, return_tensors="pt", truncation=True, max_length=512)
+    with torch.no_grad():
+        logits = detector_model(**inputs).logits
+        probs = torch.softmax(logits, dim=1).tolist()[0]
+    human_prob, ai_prob = probs[0], probs[1]
+
+    # Model 2: GPT-2 Perplexity
+    ppl = compute_perplexity(user_text)
+    # Heuristic: low perplexity → AI
+    ppl_score = max(0, min(1, 100/ppl))  # normalize to 0..1
+
+    # Aggregate
+    final_ai = (ai_prob + ppl_score) / 2
+    final_human = 1 - final_ai
+
+    return {
+        "RoBERTa AI Probability": round(ai_prob*100, 2),
+        "Perplexity-based AI Probability": round(ppl_score*100, 2),
+        "Final AI Probability (avg)": round(final_ai*100, 2),
+        "Final Human Probability (avg)": round(final_human*100, 2),
+    }
 
-# Load MAGE dataset (optional if you want examples)
-dataset = load_dataset("yaful/MAGE")
-
-# Load a pretrained AI text detector (DeBERTa Detector as example)
-model_name = "mrm8488/distilroberta-finetuned-openai-detector"
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-model = AutoModelForSequenceClassification.from_pretrained(model_name)
-classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
-
-# Function to detect Human or AI
-def detect_text(input_text):
-    if len(input_text.split()) < 3:  # Skip very short text
-        return "Text too short to classify accurately."
-    
-    result = classifier(input_text)[0]
-    label = result['label']
-    score = result['score'] * 100
-    return f"Prediction: {label}\nConfidence: {score:.2f}%"
-
-# Optional: show a sample from MAGE dataset
-def sample_text(index):
-    s = dataset['train'][index]
-    return f"Text: {s['text']}\nLabel: {s['label']}"
-
-# Gradio UI
-demo = gr.Blocks()
-
-with demo:
-    gr.Markdown("# AI vs Human Text Detector")
-    with gr.Tab("Detect Text"):
-        text_input = gr.Textbox(label="Enter Text Here", lines=5)
-        output = gr.Textbox(label="Prediction")
-        detect_btn = gr.Button("Detect")
-        detect_btn.click(detect_text, inputs=text_input, outputs=output)
-    
-    with gr.Tab("Sample Text from MAGE"):
-        index_input = gr.Number(label="Sample Index", value=0)
-        sample_output = gr.Textbox(label="Sample")
-        index_input.change(sample_text, inputs=index_input, outputs=sample_output)
+with gr.Blocks() as demo:
+    gr.Markdown("# 🔍 Free AI vs Human Text Detector (Demo)")
+    user_input = gr.Textbox(
+        label="✍️ Enter Text",
+        placeholder="Paste text here...",
+        lines=12,
+        type="text"
+    )
+    output = gr.JSON(label="Results")
+    run_btn = gr.Button("Run Detection")
+    run_btn.click(analyze_text, inputs=user_input, outputs=output)
 
 demo.launch()