app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import re
from tokenizers.normalizers import Sequence, Replace, Strip
from tokenizers import Regex
import matplotlib.pyplot as plt

# --- Setup and Model Loading ---
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Model paths and URLs
model1_path = "modernbert.bin"
model2_url = "https://huggingface.co/mihalykiss/modernbert_2/resolve/main/Model_groups_3class_seed12"
model3_url = "https://huggingface.co/mihalykiss/modernbert_2/resolve/main/Model_groups_3class_seed22"

tokenizer = AutoTokenizer.from_pretrained("answerdotai/ModernBERT-base")

def load_model(path_or_url, num_labels=41):
    model = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=num_labels)
    if path_or_url.startswith("http"):
        # Load from Hugging Face URL
        state_dict = torch.hub.load_state_dict_from_url(path_or_url, map_location=device)
    else:
        # Load from local file
        state_dict = torch.load(path_or_url, map_location=device)
    model.load_state_dict(state_dict)
    return model.to(device).eval()

# Initializing the ensemble
model_1 = load_model(model1_path)
model_2 = load_model(model2_url)
model_3 = load_model(model3_url)

label_mapping = {
    0: '13B', 1: '30B', 2: '65B', 3: '7B', 4: 'GLM130B', 5: 'bloom_7b',
    6: 'bloomz', 7: 'cohere', 8: 'davinci', 9: 'dolly', 10: 'dolly-v2-12b',
    11: 'flan_t5_base', 12: 'flan_t5_large', 13: 'flan_t5_small',
    14: 'flan_t5_xl', 15: 'flan_t5_xxl', 16: 'gemma-7b-it', 17: 'gemma2-9b-it',
    18: 'gpt-3.5-turbo', 19: 'gpt-35', 20: 'gpt4', 21: 'gpt4o',
    22: 'gpt_j', 23: 'gpt_neox', 24: 'human', 25: 'llama3-70b', 26: 'llama3-8b',
    27: 'mixtral-8x7b', 28: 'opt_1.3b', 29: 'opt_125m', 30: 'opt_13b',
    31: 'opt_2.7b', 32: 'opt_30b', 33: 'opt_350m', 34: 'opt_6.7b',
    35: 'opt_iml_30b', 36: 'opt_iml_max_1.3b', 37: 't0_11b', 38: 't0_3b',
    39: 'text-davinci-002', 40: 'text-davinci-003'
}

# --- Text Preprocessing ---
def clean_text(text: str) -> str:
    text = re.sub(r'\s{2,}', ' ', text)
    text = re.sub(r'\s+([,.;:?!])', r'\1', text)
    return text

# Custom Tokenizer Normalization
newline_to_space = Replace(Regex(r'\s*\n\s*'), " ")
join_hyphen_break = Replace(Regex(r'(\w+)[--]\s*\n\s*(\w+)'), r"\1\2")
tokenizer.backend_tokenizer.normalizer = Sequence([
    tokenizer.backend_tokenizer.normalizer,
    join_hyphen_break,
    newline_to_space,
    Strip()
])

# --- Core Classification Logic ---
def classify_text(text):
    cleaned_text = clean_text(text)
    if not cleaned_text.strip():
        return "Please enter text to analyze.", None

    inputs = tokenizer(cleaned_text, return_tensors="pt", truncation=True, padding=True).to(device)

    with torch.no_grad():
        logits_1 = model_1(**inputs).logits
        logits_2 = model_2(**inputs).logits
        logits_3 = model_3(**inputs).logits

        # Soft voting ensemble (averaging probabilities)
        s1, s2, s3 = torch.softmax(logits_1, 1), torch.softmax(logits_2, 1), torch.softmax(logits_3, 1)
        avg_probs = (s1 + s2 + s3) / 3
        probs = avg_probs[0]

    # Calculate probabilities for the 2 main categories (Human vs AI)
    human_prob = probs[24].item()
    
    # To find the specific LLM, we ignore the 'human' index (24)
    ai_probs_only = probs.clone()
    ai_probs_only[24] = 0 
    ai_total_prob = ai_probs_only.sum().item()

    # Normalize percentages
    total_sum = human_prob + ai_total_prob
    human_pct = (human_prob / total_sum) * 100
    ai_pct = (ai_total_prob / total_sum) * 100

    # Identify the specific AI model with the highest sub-probability
    top_ai_idx = torch.argmax(ai_probs_only).item()
    predicted_llm = label_mapping[top_ai_idx]

    # Construct the result display
    if human_pct > ai_pct:
        result_message = (
            f"### Result: <span class='highlight-human'>**{human_pct:.2f}% Human written**</span>\n\n"
            "The content matches human writing patterns."
        )
    else:
        result_message = (
            f"### Result: <span class='highlight-ai'>**{ai_pct:.2f}% AI generated**</span>\n\n"
            f"**Specific Model Identified:** `{predicted_llm}`\n\n"
            "The structure and syntax are highly characteristic of this LLM."
        )

    # Visualization
    fig, ax = plt.subplots(figsize=(8, 4))
    bars = ax.bar(['Human', 'AI'], [human_pct, ai_pct], color=['#4CAF50', '#FF5733'], alpha=0.8)
    ax.set_ylabel('Probability (%)')
    ax.set_title('Detection Probability')
    ax.set_ylim(0, 110) # Room for text labels
    
    for bar in bars:
        height = bar.get_height()
        ax.text(bar.get_x() + bar.get_width()/2., height + 2, f'{height:.1f}%', ha='center', fontweight='bold')
    
    plt.tight_layout()
    return result_message, fig

# --- Gradio UI Layout ---
with gr.Blocks(css="""
    .highlight-human { color: #4CAF50; font-weight: bold; background: rgba(76, 175, 80, 0.1); padding: 5px; border-radius: 5px; }
    .highlight-ai { color: #FF5733; font-weight: bold; background: rgba(255, 87, 51, 0.1); padding: 5px; border-radius: 5px; }
    #output-container { text-align: center; padding: 20px; }
""") as iface:
    gr.Markdown("# AI Text Detector & LLM Identifier")
    gr.Markdown("This tool uses an ensemble of **ModernBERT** models to predict if text is human or AI, and specifies the likely source model.")
    
    with gr.Row():
        with gr.Column(scale=2):
            text_input = gr.Textbox(
                label="Input Text", 
                placeholder="Paste your English text here...", 
                lines=10
            )
        with gr.Column(scale=1):
            result_output = gr.Markdown("Analysis results will appear here.", elem_id="output-container")
            plot_output = gr.Plot()

    # Trigger classification on text change
    text_input.change(classify_text, inputs=text_input, outputs=[result_output, plot_output])

    gr.Markdown("---")
    gr.Markdown("**Developed by SzegedAI**")

if __name__ == "__main__":
    # share=True creates a public URL
    iface.launch(share=True)