app.py

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

# ----------------------
# Paraphrasing Model Setup (Pegasus)
# ----------------------
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)

# ----------------------
# Semantic Similarity Model
# ----------------------
similarity_model = SentenceTransformer('all-MiniLM-L6-v2')

# ----------------------
# Local AI Detector (roberta-base-openai-detector)
# ----------------------
AI_DETECTOR_MODEL = "roberta-base-openai-detector"
ai_detector = pipeline("text-classification", model=AI_DETECTOR_MODEL, device=0 if torch.cuda.is_available() else -1)

# ----------------------
# Prompt Variations for Humanization
# ----------------------
PROMPT_VARIANTS = [
    "Paraphrase this naturally:",
    "Rewrite as if explaining to a friend:",
    "Make this sound like a real conversation:",
    "Express this in a casual, human way:",
    "Reword this with natural flow:",
    "Make this sound less robotic:",
    "Rewrite in a friendly, informal tone:",
    "Paraphrase in a way a student would say it:",
]

# ----------------------
# Sentence Splitter
# ----------------------
def split_sentences(text):
    sentences = re.split(r'(?<=[.!?])\s+', text.strip())
    return [s for s in sentences if s]

# ----------------------
# Light Post-Processing
# ----------------------
def postprocess_text(text):
    contractions = {
        "do not": "don't", "cannot": "can't", "will not": "won't", "I am": "I'm",
        "is not": "isn't", "are not": "aren't", "did not": "didn't", "it is": "it's",
        "does not": "doesn't", "have not": "haven't", "has not": "hasn't"
    }
    for k, v in contractions.items():
        text = re.sub(rf'\b{k}\b', v, text, flags=re.IGNORECASE)
    idioms = [
        "at the end of the day", "to be honest", "as a matter of fact", "for what it's worth",
        "in a nutshell", "the bottom line is", "all things considered"
    ]
    if random.random() < 0.3:
        text += " " + random.choice(idioms) + "."
    return text

# ----------------------
# Sentence-level Paraphrasing with Prompt Variation
# ----------------------
def paraphrase_sentence(sentence, tone):
    prompt = random.choice(PROMPT_VARIANTS)
    if tone != "Stealth":
        prompt = f"{prompt} ({tone} tone):"
    full_prompt = f"{prompt} {sentence}"
    batch = paraphrase_tokenizer([full_prompt], truncation=True, padding='longest', max_length=60, return_tensors="pt").to(device)
    outputs = paraphrase_model.generate(
        **batch,
        max_length=60,
        num_beams=5,
        num_return_sequences=1,
        temperature=1.0
    )
    tgt_text = paraphrase_tokenizer.batch_decode(outputs, skip_special_tokens=True)
    return tgt_text[0] if tgt_text else sentence

# ----------------------
# Main Paraphrasing Function
# ----------------------
def paraphrase(text, tone):
    sentences = split_sentences(text)
    paraphrased = []
    for sent in sentences:
        rewritten = paraphrase_sentence(sent, tone)
        paraphrased.append(rewritten)
    joined = ' '.join(paraphrased)
    return postprocess_text(joined)

# ----------------------
# Semantic Similarity Function
# ----------------------
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

# ----------------------
# Local AI Detection Function
# ----------------------
def check_ai_score(text):
    try:
        result = ai_detector(text)
        for r in result:
            # LABEL_1 = AI, LABEL_0 = Human
            if r['label'] in ['LABEL_1', 'Fake']:
                return r['score'], None
            elif r['label'] in ['LABEL_0', 'Real']:
                return 1.0 - r['score'], None
        return 0.5, None  # fallback
    except Exception as e:
        return None, f"AI detection error: {str(e)}"

# ----------------------
# Humanization Score & Rating
# ----------------------
def humanization_score(sim, ai_prob):
    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 AI-like ({score:.2f})"
    elif score < 0.85:
        return f"👍 Acceptable ({score:.2f})"
    else:
        return f"✅ Highly Humanized ({score:.2f})"

# ----------------------
# Main Processing Function
# ----------------------
def process(text, tone):
    if not text.strip():
        return "", "", 0.0, "", 0.0, ""
    pre_ai_prob, pre_err = check_ai_score(text)
    if pre_ai_prob is None:
        return "", f"AI Detection Error: {pre_err}", 0.0, "", 0.0, ""
    try:
        paraphrased = paraphrase(text, tone)
    except Exception as e:
        return f"[Paraphrasing error: {str(e)}]", "", 0.0, "", 0.0, ""
    post_ai_prob, post_err = check_ai_score(paraphrased)
    if post_ai_prob is None:
        return paraphrased, f"AI Detection Error: {post_err}", 0.0, "", 0.0, ""
    sim = semantic_similarity(text, paraphrased)
    score = humanization_score(sim, post_ai_prob)
    rating = humanization_rating(score)
    ai_score_str = f"Pre: {100*(1-pre_ai_prob):.1f}% human | Post: {100*(1-post_ai_prob):.1f}% human"
    return (
        paraphrased,
        ai_score_str,
        sim,
        rating,
        score * 100,
        ""
    )

# ----------------------
# Gradio UI
# ----------------------
custom_theme = grthemes.Base(
    primary_hue="blue",
    secondary_hue="blue",
    neutral_hue="slate"
)

with gr.Blocks(theme=custom_theme, title="AI Humanizer - Made by Taha") 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...", elem_id="input-box")
            tone = gr.Dropdown(["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, elem_id="output-box")
            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, gr.Textbox(visible=False)],
        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 and creators
    </div>
    """, elem_id="footer")

demo.launch()