app.py

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


model1_path = "https://huggingface.co/spaces/SzegedAI/AI_Detector/resolve/main/modernbert.bin"
model2_path = "https://huggingface.co/mihalykiss/modernbert_2/resolve/main/Model_groups_3class_seed12"
model3_path = "https://huggingface.co/mihalykiss/modernbert_2/resolve/main/Model_groups_3class_seed22"

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")

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

    model_1 = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=41)
    model_1.load_state_dict(torch.hub.load_state_dict_from_url(model1_path, map_location=device, progress=True))
    model_1.to(device).eval()

    model_2 = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=41)
    model_2.load_state_dict(torch.hub.load_state_dict_from_url(model2_path, map_location=device, progress=True))
    model_2.to(device).eval()

    model_3 = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=41)
    model_3.load_state_dict(torch.hub.load_state_dict_from_url(model3_path, map_location=device, progress=True))
    model_3.to(device).eval()

except Exception as e:
    print(f"Error during model loading: {e}")
    print("Please ensure all model paths are correct, dependencies are installed, and you have an internet connection for remote models.")
    # Handle the error, e.g., by exiting or displaying an error in the UI if Gradio is already running.
    # For simplicity, we'll let it potentially crash if models can't load before Gradio starts.
    # If Gradio is already running, you'd need a more sophisticated error display.
    # For now, we'll just make sure the Gradio interface doesn't try to use non-existent models.
    tokenizer = None # Prevent further errors if tokenizer failed
    model_1, model_2, model_3 = None, None, None


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'
}

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

if tokenizer: # Only set normalizer if tokenizer loaded successfully
    newline_to_space = Replace(Regex(r'\s*\n\s*'), " ")
    join_hyphen_break = Replace(Regex(r'(\w+)[--]\s*\n\s*(\w+)'), r"\1\2") # Corrected hyphen regex
    tokenizer.backend_tokenizer.normalizer = Sequence([
        tokenizer.backend_tokenizer.normalizer, # Keep existing normalizers
        join_hyphen_break,
        newline_to_space,
        Strip()
    ])
# --- End Model & Tokenizer Configuration ---


title_md = """
<h1 style="text-align: center; margin-bottom: 5px;">AI Text Detector</h1>
<p style="text-align: center; font-size: 0.9em; color: var(--text-secondary); margin-top: 0; margin-bottom: 20px;">Developed by SzegedAI</p>
"""

description = """
<div class="app-description">
    <p>This tool utilizes the <b>ModernBERT</b> model to discern whether a given text is human-authored or AI-generated. It employs a soft voting ensemble of <b>three</b> models, amalgamating their outputs to enhance detection accuracy.</p>
    <ul class="features-list">
        <li><span class="icon">✅</span> <strong>Human Verification:</strong> Clearly identifies human-written content.</li>
        <li><span class="icon">🔍</span> <strong>Model Detection:</strong> Capable of identifying content from over 40 AI models.</li>
        <li><span class="icon">📈</span> <strong>Accuracy:</strong> Performs optimally with more extensive text inputs.</li>
        <li><span class="icon">📄</span> <strong>Read more:</strong> Our methodology is detailed in our research paper:
            <a href="https://aclanthology.org/2025.genaidetect-1.15/" target="_blank" class="learn-more-link"><b> LINK</b></a>.
        </li>
    </ul>
    <p class="instruction-text">Paste your text into the field below to analyze its origin.</p>
</div>
"""
bottom_text = "<p class='footer-text'>SzegedAI</p>" # Simplified footer, as requested

AI_texts = [
    "Camels are remarkable desert animals known for their unique adaptations to harsh, arid environments. Native to the Middle East, North Africa, and parts of Asia, camels have been essential to human life for centuries, serving as a mode of transportation, a source of food, and even a symbol of endurance and survival. There are two primary species of camels: the dromedary camel, which has a single hump and is commonly found in the Middle East and North Africa, and the Bactrian camel, which has two humps and is native to Central Asia. Their humps store fat, not water, as commonly believed, allowing them to survive long periods without food by metabolizing the stored fat for energy. Camels are highly adapted to desert life. They can go for weeks without water, and when they do drink, they can consume up to 40 gallons in one sitting. Their thick eyelashes, sealable nostrils, and wide, padded feet protect them from sand and help them walk easily on loose desert terrain.",
]

Human_texts = [
    "To make BERT handle a variety of down-stream tasks, our input representation is able to unambiguously represent both a single sentence and a pair of sentences (e.g., h Question, Answeri) in one token sequence. Throughout this work, a “sentence” can be an arbitrary span of contiguous text, rather than an actual linguistic sentence. A “sequence” refers to the input token sequence to BERT, which may be a single sentence or two sentences packed together. We use WordPiece embeddings (Wu et al., 2016) with a 30,000 token vocabulary. The first token of every sequence is always a special classification token ([CLS]). The final hidden state corresponding to this token is used as the aggregate sequence representation for classification tasks. Sentence pairs are packed together into a single sequence."
]

def classify_text_interface(text):
    if not all([tokenizer, model_1, model_2, model_3]):
        return "<p style='text-align: center; color: var(--ai-color);'><strong>Error: Models not loaded. Please check the console.</strong></p>"

    cleaned_text = clean_text(text)
    if not cleaned_text.strip(): # Check cleaned_text here
        result_message = "<p style='text-align: center; color: var(--text-secondary);'>Please enter some text to analyze.</p>"
        return result_message

    inputs = tokenizer(cleaned_text, return_tensors="pt", truncation=True, padding=True, max_length=512).to(device) # Added max_length
    
    with torch.no_grad():
        logits_1 = model_1(**inputs).logits
        logits_2 = model_2(**inputs).logits
        logits_3 = model_3(**inputs).logits
        
        softmax_1 = torch.softmax(logits_1, dim=1)
        softmax_2 = torch.softmax(logits_2, dim=1)
        softmax_3 = torch.softmax(logits_3, dim=1)
        
        averaged_probabilities = (softmax_1 + softmax_2 + softmax_3) / 3
        probabilities = averaged_probabilities[0]

    ai_probs = probabilities.clone()
    human_label_index = -1
    for k, v in label_mapping.items(): # Find the human label index dynamically
        if v.lower() == 'human':
            human_label_index = k
            break
    
    if human_label_index != -1:
        ai_probs[human_label_index] = 0 # Zero out human probability for AI sum
        human_prob_value = probabilities[human_label_index].item() * 100
    else: # Fallback if 'human' not in label_mapping (should not happen with current map)
        human_prob_value = 0 
        print("Warning: 'human' label not found in label_mapping.")

    ai_total_prob = ai_probs.sum().item() * 100

    # Recalculate human_prob based on ai_total_prob if necessary,
    # or ensure the logic correctly identifies human vs AI majority.
    # The original logic: human_prob = 100 - ai_total_prob might be confusing if ai_total_prob already excluded human.
    # Let's use the direct human probability from the model.
    
    ai_argmax_index = torch.argmax(ai_probs).item() # Argmax over non-human probabilities
    ai_argmax_model = label_mapping.get(ai_argmax_index, "Unknown AI")

    if human_prob_value > ai_total_prob : # Compare direct human probability with sum of AI probabilities
        result_message = (
            f"<p><strong>The text is</strong> <span class='highlight-human'><strong>{human_prob_value:.2f}%</strong> likely <b>Human written</b>.</span></p>"
        )
    else:
        result_message = (
            f"<p><strong>The text is</strong> <span class='highlight-ai'><strong>{ai_total_prob:.2f}%</strong> likely <b>AI generated</b>.</span></p>"
            f"<p style='margin-top: 10px; font-size: 0.95em;'><strong>Most Likely AI Source:</strong> {ai_argmax_model} (with {probabilities[ai_argmax_index].item()*100:.2f}% confidence among AI models)</p>"
        )
    return result_message

modern_css = """
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap');

:root {
    --primary-bg: #F8F9FA;
    --app-bg: #FFFFFF;
    --text-primary: #2C3E50;
    --text-secondary: #7F8C8D;
    --accent-color: #1ABC9C;
    --accent-color-darker: #16A085;
    --border-color: #E0E0E0;
    --input-bg: #FFFFFF;
    --input-focus-border: var(--accent-color);
    --human-color: #2ECC71; /* Green */
    --human-bg: rgba(46, 204, 113, 0.1);
    --ai-color: #E74C3C; /* Red */
    --ai-bg: rgba(231, 76, 60, 0.1);
    --shadow-color: rgba(44, 62, 80, 0.1);
    --container-max-width: 800px; /* Increased width */
    --border-radius-md: 8px;
    --border-radius-lg: 12px;
}

body {
    font-family: 'Inter', sans-serif;
    background: linear-gradient(135deg, #f5f7fa 0%, #eef2f7 100%);
    color: var(--text-primary);
    margin: 0;
    padding: 20px;
    display: flex;
    justify-content: center;
    align-items: flex-start;
    min-height: 100vh;
    box-sizing: border-box;
    overflow-y: auto;
}

.gradio-container {
    background-color: var(--app-bg);
    border-radius: var(--border-radius-lg);
    padding: clamp(25px, 5vw, 40px);
    box-shadow: 0 8px 25px var(--shadow-color);
    max-width: var(--container-max-width);
    width: 100%;
    margin: 20px auto;
    border: none;
}

.form.svelte-633qhp, .block.svelte-11xb1hd, .gradio-html .block { /* More generic selector for Gradio HTML block */
    background: none !important;
    border: none !important;
    box-shadow: none !important;
    padding: 0 !important;
}

/* Title and subtitle are now handled by Markdown with inline styles, h1 here is a fallback or for other h1s */
h1 { 
    color: var(--text-primary);
    font-size: clamp(24px, 5vw, 30px); 
    font-weight: 700;
    text-align: center;
    margin-bottom: 20px; /* Adjusted default h1 margin */
    letter-spacing: -0.5px;
}

.app-description p {
    color: var(--text-secondary);
    font-size: clamp(14px, 2.5vw, 16px);
    line-height: 1.7;
    margin-bottom: 15px;
}

.app-description .instruction-text {
    font-weight: 500;
    color: var(--text-primary);
    margin-top: 20px;
    text-align: center;
}

.features-list {
    list-style: none;
    padding-left: 0;
    margin: 20px 0;
}

.features-list li {
    display: flex;
    align-items: center;
    font-size: clamp(14px, 2.5vw, 16px);
    color: var(--text-secondary);
    margin-bottom: 12px;
    line-height: 1.6;
}

.features-list .icon {
    margin-right: 12px;
    font-size: 1.2em;
    color: var(--accent-color);
}

.learn-more-link, .learn-more-link b {
    color: var(--accent-color) !important;
    text-decoration: none;
    font-weight: 600;
}
.learn-more-link:hover, .learn-more-link:hover b {
    color: var(--accent-color-darker) !important;
    text-decoration: underline;
}

#text_input_box textarea {
    background-color: var(--input-bg);
    border: 1px solid var(--border-color);
    border-radius: var(--border-radius-md);
    font-size: clamp(15px, 2.5vw, 16px);
    padding: 15px;
    width: 100%;
    box-sizing: border-box;
    color: var(--text-primary);
    transition: border-color 0.3s ease, box-shadow 0.3s ease;
    min-height: 120px;
    box-shadow: 0 2px 4px rgba(0,0,0,0.05);
}

#text_input_box textarea::placeholder {
    color: #B0BEC5;
}

#text_input_box textarea:focus {
    border-color: var(--input-focus-border);
    box-shadow: 0 0 0 3px rgba(26, 188, 156, 0.2);
    outline: none;
}

#result_output_box { 
    background-color: var(--input-bg); /* Ensure background for the box */
    border: 1px solid var(--border-color);
    border-radius: var(--border-radius-md);
    padding: 20px;
    margin-top: 25px;
    width: 100%;
    box-sizing: border-box;
    text-align: center;
    font-size: clamp(16px, 3vw, 17px); /* Slightly adjusted font size for results */
    box-shadow: 0 4px 8px rgba(0,0,0,0.05);
    min-height: 80px; /* Give it some min height */
    display: flex; /* For centering content if needed */
    flex-direction: column;
    justify-content: center;
}
#result_output_box p { /* Style paragraphs inside the result box */
    margin-bottom: 8px; /* Space between lines in result */
    line-height: 1.6;
}
#result_output_box p:last-child {
    margin-bottom: 0;
}


.highlight-human, .highlight-ai {
    font-weight: 600;
    padding: 5px 10px; /* Adjusted padding */
    border-radius: var(--border-radius-md);
    display: inline-block;
    font-size: 1.05em; /* Adjusted size */
}

.highlight-human {
    color: var(--human-color);
    background-color: var(--human-bg);
    /* border: 1px solid var(--human-color); Removed border for cleaner look */
}

.highlight-ai {
    color: var(--ai-color);
    background-color: var(--ai-bg);
    /* border: 1px solid var(--ai-color); Removed border for cleaner look */
}

.tabs > div:first-child button {
    background-color: transparent !important;
    color: var(--text-secondary) !important;
    border: none !important;
    border-bottom: 2px solid transparent !important;
    border-radius: 0 !important;
    padding: 10px 15px !important;
    font-weight: 500 !important;
    transition: color 0.3s ease, border-bottom-color 0.3s ease !important;
}

.tabs > div:first-child button.selected {
    color: var(--accent-color) !important;
    border-bottom-color: var(--accent-color) !important;
    font-weight: 600 !important;
}

.gr-examples {
    padding: 15px !important;
    border: 1px solid var(--border-color) !important;
    border-radius: var(--border-radius-md) !important;
    background-color: #fdfdfd !important;
    margin-top: 10px; /* Add some space above examples */
}
.gr-sample-textbox {
    border: 1px solid var(--border-color) !important;
    border-radius: var(--border-radius-md) !important;
    font-size: 14px !important;
}
.gr-accordion > .label-wrap button { /* Style accordion label */
    font-weight: 500 !important;
    color: var(--text-primary) !important;
}


.footer-text, #bottom_text {
    text-align: center;
    margin-top: 40px;
    font-size: clamp(13px, 2vw, 14px);
    color: var(--text-secondary);
}
#bottom_text p { 
    margin: 0;
}

@media (max-width: 768px) {
    body {
        padding: 10px;
        align-items: flex-start;
    }
    .gradio-container {
        padding: 20px;
        margin: 10px;
    }
    h1 { font-size: 22px; } /* Adjust for custom title markdown */
    .app-description p, .features-list li { font-size: 14px; }
    #text_input_box textarea { font-size: 15px; min-height: 100px; }
    #result_output_box { font-size: 15px; padding: 15px; }
}
"""

iface = gr.Blocks(css=modern_css, theme=gr.themes.Base(font=[gr.themes.GoogleFont("Inter"), "sans-serif"]))

with iface:
    gr.Markdown(title_md) # Using combined Markdown for title and subtitle
    gr.Markdown(description)

    text_input = gr.Textbox(
        label="",
        placeholder="Type or paste your content here...",
        elem_id="text_input_box",
        lines=7 # Adjusted lines
    )
    result_output = gr.HTML(elem_id="result_output_box")

    # Only set up the change function if models are loaded
    if all([tokenizer, model_1, model_2, model_3]):
        text_input.change(classify_text_interface, inputs=text_input, outputs=result_output)
    else:
        # Display a persistent error if models couldn't load
        gr.HTML("<div id='result_output_box'><p style='color: var(--ai-color); text-align: center;'><strong>Application Error: Models could not be loaded. Please check the server console for details.</strong></p></div>")


    with gr.Row():
        with gr.Column(scale=1):
             with gr.Accordion("AI Text Examples", open=False):
                gr.Examples(
                    examples=AI_texts,
                    inputs=text_input,
                    label="", # Label removed as accordion title is enough
                )
        with gr.Column(scale=1):
            with gr.Accordion("Human Text Examples", open=False):
                gr.Examples(
                    examples=Human_texts,
                    inputs=text_input,
                    label="", # Label removed
                )

    gr.Markdown(bottom_text, elem_id="bottom_text")

if __name__ == "__main__":
    iface.launch(share=False)