streamlit_app.py

import streamlit as st
import joblib
import torch
from transformers import GPT2LMHeadModel, GPT2TokenizerFast
import numpy as np

# -------------------------------
# Load Logistic Regression model
# -------------------------------
vectorizer = joblib.load('src/vectorizer.pkl')
model = joblib.load('src/logistic_model.pkl')

# -------------------------------
# Load GPT-2 model and tokenizer
# -------------------------------
@st.cache_resource
def load_gpt2():
    tokenizer = GPT2TokenizerFast.from_pretrained("src/gpt2_local")
    gpt2_model = GPT2LMHeadModel.from_pretrained("src/gpt2_local")
    gpt2_model.eval()
    return tokenizer, gpt2_model

tokenizer, gpt2_model = load_gpt2()

# -------------------------------
# Calculate Perplexity
# -------------------------------
def calculate_perplexity(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        outputs = gpt2_model(**inputs, labels=inputs["input_ids"])
        loss = outputs.loss
        perplexity = torch.exp(loss).item()
    return perplexity

# -------------------------------
# Final prediction combining both
# -------------------------------
def final_score(ai_prob, perplexity):
    # Normalize perplexity: higher => more human-like (score 0), lower => more AI-like (score 1)
    if perplexity > 300:
        perp_score = 0.0
    elif perplexity < 10:
        perp_score = 1.0
    else:
        perp_score = 1.0 - ((perplexity - 10) / (300 - 10))  # scale to [0,1]
        perp_score = max(0.0, min(1.0, perp_score))

    # Combine: 70% weight to perplexity, 30% to logistic regression
    final_ai_score = (0.7 * perp_score) + (0.3 * ai_prob)
    return final_ai_score, perp_score

# -------------------------------
# Streamlit UI
# -------------------------------
st.set_page_config(page_title="AI Text Detector", page_icon="🤖", layout="centered")

st.markdown("""
<h2 style='text-align: center; color: #4CAF50;'>🤖 AI vs Human Text Detector</h2>
<p style='text-align: center;'>Enter a sentence to check if it was written by a human or generated by AI.</p>
""", unsafe_allow_html=True)

user_input = st.text_area("Enter your sentence here:", height=150)

if st.button("Check"):
    if user_input.strip() == "":
        st.warning("Please enter a sentence before submitting.")
    else:
        # Logistic model prediction
        transformed_input = vectorizer.transform([user_input])
        prediction = model.predict_proba(transformed_input)
        ai_prob = prediction[0][1]
        human_prob = prediction[0][0]

        # Perplexity calculation
        perplexity_score = calculate_perplexity(user_input)

        # Combine scores
        final_ai_score, perp_score = final_score(ai_prob, perplexity_score)

        # Display Results
        st.subheader("🔍 Result:")
        if final_ai_score > 0.5:
            st.error("❗ This text is likely **AI-generated**.")
        else:
            st.success("✅ This text is likely **Human-written**.")

        st.markdown(f"**Logistic Model Confidence:** {ai_prob:.3f} AI vs {human_prob:.3f} Human")
        st.markdown(f"**Perplexity Score:** {perplexity_score:.2f}")
        st.markdown(f"**Combined AI Score:** {final_ai_score:.3f} (Weighted)")

        # Interpretation
        if perplexity_score < 30:
            st.info("🧠 Low perplexity suggests the text is highly predictable—possibly AI-generated.")
        elif perplexity_score > 100:
            st.info("🧠 High perplexity suggests human-like variation or complexity.")