app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
from transformers.pipelines import pipeline
from sentence_transformers import SentenceTransformer, util
import numpy as np
import gradio.themes as grthemes

# Paraphrasing model: tuner007/pegasus_paraphrase
PARAPHRASE_MODEL_NAME = "tuner007/pegasus_paraphrase"
paraphrase_tokenizer = AutoTokenizer.from_pretrained(PARAPHRASE_MODEL_NAME)
paraphrase_model = AutoModelForSeq2SeqLM.from_pretrained(PARAPHRASE_MODEL_NAME)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
paraphrase_model = paraphrase_model.to(device)

# AI Detector: desklib/ai-text-detector-v1.01
AI_DETECTOR_MODEL = "desklib/ai-text-detector-v1.01"
ai_detector = pipeline("text-classification", model=AI_DETECTOR_MODEL, device=0 if torch.cuda.is_available() else -1)

# Semantic similarity model
similarity_model = SentenceTransformer('all-MiniLM-L6-v2')

def paraphrase(text):
    prompt = text.strip()
    batch = paraphrase_tokenizer([prompt], truncation=True, padding='longest', max_length=60, return_tensors="pt").to(device)
    translated = paraphrase_model.generate(
        **batch,
        max_length=60,
        num_beams=5,
        num_return_sequences=1,
        temperature=1.0
    )
    tgt_text = paraphrase_tokenizer.batch_decode(translated, skip_special_tokens=True)
    return tgt_text[0] if tgt_text else ""

def semantic_similarity(text1, text2):
    emb1 = similarity_model.encode(text1, convert_to_tensor=True)
    emb2 = similarity_model.encode(text2, convert_to_tensor=True)
    sim = util.pytorch_cos_sim(emb1, emb2).item()
    return sim

def ai_detect(text):
    # Returns probability of being AI-generated (label 'LABEL_1' = AI, 'LABEL_0' = Human)
    result = ai_detector(text)
    for r in result:
        if r['label'] in ['LABEL_1', 'Fake']:
            return r['score']
        elif r['label'] in ['LABEL_0', 'Real']:
            return 1.0 - r['score']
    return 0.5  # fallback

def humanization_score(sim, ai_prob):
    # Lower similarity and lower AI probability = more human
    score = (1.0 - sim) * 0.5 + (1.0 - ai_prob) * 0.5
    return score

def humanization_rating(score):
    if score < 0.7:
        return f"⚠️ Still robotic ({score:.2f})"
    elif score < 0.85:
        return f"👍 Acceptable ({score:.2f})"
    else:
        return f"✅ Highly Human ({score:.2f})"

def process(text, tone):
    if not text.strip():
        return "", "", 0.0, "", 0.0
    # Pre-humanization AI detection
    pre_ai_prob = ai_detect(text)
    # Paraphrase
    try:
        paraphrased = paraphrase(text)
    except Exception as e:
        return "[Error in paraphrasing: {}]".format(str(e)), "", 0.0, "", 0.0
    # Post-humanization AI detection
    post_ai_prob = ai_detect(paraphrased)
    # Semantic similarity
    sim = semantic_similarity(text, paraphrased)
    # Humanization score
    score = humanization_score(sim, post_ai_prob)
    rating = humanization_rating(score)
    ai_score_str = f"Pre: {pre_ai_prob*100:.1f}% | Post: {post_ai_prob*100:.1f}%"
    return (
        paraphrased,         # gr.Textbox (string)
        ai_score_str,        # gr.Markdown (string)
        sim,                # gr.Number (float)
        rating,             # gr.Markdown (string)
        score * 100         # gr.Number (float)
    )

# Custom dark theme using gradio.themes.Base
custom_theme = grthemes.Base(
    primary_hue="blue",
    secondary_hue="blue",
    neutral_hue="slate"
)

with gr.Blocks(theme=custom_theme) as demo:
    gr.Markdown("""
    # 🧠 AI Humanizer
    <div style='display:flex;justify-content:space-between;align-items:center;'>
        <span style='font-size:1.2em;color:#7bb1ff;'>Rewrite AI text to sound 100% human</span>
        <span style='font-weight:bold;color:#7bb1ff;'>Made by Taha</span>
    </div>
    """, elem_id="header")
    with gr.Row():
        with gr.Column():
            text_in = gr.Textbox(label="Paste AI-generated text here", lines=8, placeholder="Paste your text...")
            tone = gr.Radio(["Academic", "Casual", "Friendly", "Stealth"], value="Stealth", label="Tone Selector")
            btn = gr.Button("Humanize", elem_id="humanize-btn")
        with gr.Column():
            text_out = gr.Textbox(label="Humanized Output", lines=8, interactive=False)
            ai_scores = gr.Markdown("", elem_id="ai-scores")
            sim_score = gr.Number(label="Similarity (0=very different, 1=very similar)", interactive=False)
            rating = gr.Markdown("", elem_id="rating")
            human_score = gr.Number(label="Humanization Score (%)", interactive=False)
    btn.click(
        process,
        inputs=[text_in, tone],
        outputs=[text_out, ai_scores, sim_score, rating, human_score],
        api_name="humanize"
    )
    gr.Markdown("""
    <div style='text-align:center;color:#7bb1ff;margin-top:2em;'>
        <b>Made by Taha</b> | Free for unlimited use | Optimized for students
    </div>
    """, elem_id="footer")

demo.launch()