also print the possibility of which model is used

app.py

@@ -2,31 +2,35 @@ import gradio as gr
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 import torch
 import re
-from tokenizers import normalizers
-from tokenizers.normalizers import Sequence, Replace, Strip, NFKC
+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_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"
+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")
 
-model_1 = AutoModelForSequenceClassification.from_pretrained("answerdotai/ModernBERT-base", num_labels=41)
-model_1.load_state_dict(torch.load(model1_path, map_location=device))
-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))
-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))
-model_3.to(device).eval()
+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',
@@ -41,11 +45,13 @@ label_mapping = {
     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([
@@ -55,15 +61,11 @@ tokenizer.backend_tokenizer.normalizer = Sequence([
     Strip()
 ])
 
-
+# --- Core Classification Logic ---
 def classify_text(text):
-    """
-    Classifies the text and generates a plot of the human vs AI probability.
-    Returns both the result message and the plot figure.
-    """
     cleaned_text = clean_text(text)
     if not cleaned_text.strip():
-        return "", None
+        return "Please enter text to analyze.", None
 
     inputs = tokenizer(cleaned_text, return_tensors="pt", truncation=True, padding=True).to(device)
 
@@ -72,142 +74,81 @@ def classify_text(text):
         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]
-
-    human_prob = probabilities[24].item()
-    ai_probs_clone = probabilities.clone()
-    ai_probs_clone[24] = 0
-    ai_total_prob = ai_probs_clone.sum().item()
-
-    total_decision_prob = human_prob + ai_total_prob
-    human_percentage = (human_prob / total_decision_prob) * 100
-    ai_percentage = (ai_total_prob / total_decision_prob) * 100
-
-    ai_argmax_index = torch.argmax(ai_probs_clone).item()
-    ai_argmax_model = label_mapping[ai_argmax_index]
-
-    if human_percentage > ai_percentage:
+        # 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"**The text is** <span class='highlight-human'>**{human_percentage:.2f}%** likely <b>Human written</b>.</span>"
+            f"### Result: <span class='highlight-human'>**{human_pct:.2f}% Human written**</span>\n\n"
+            "The content matches human writing patterns."
         )
     else:
         result_message = (
-            f"**The text is** <span class='highlight-ai'>**{ai_percentage:.2f}%** likely <b>AI generated</b>.</span>\n\n"
+            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."
         )
 
-    fig, ax = plt.subplots(figsize=(8, 4))  # Adjust figure size for better layout
-
-    categories = ['Human', 'AI']
-    probabilities_for_plot = [human_percentage, ai_percentage]
-
-    bars = ax.bar(categories, probabilities_for_plot, color=['#4CAF50', '#FF5733'], alpha=0.8)
-    ax.set_ylabel('Probability (%)', fontsize=12)
-    ax.set_title('Human vs AI Probability', fontsize=14, fontweight='bold')
-    ax.grid(axis='y', linestyle='--', alpha=0.6)
-
-    # Add labels to the bars
+    # 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 + 1, f'{height:.2f}%', ha='center')
-
-    ax.set_ylim(0, 100) 
+        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
 
-
-
-
-
-title = "AI Text Detector"
-
-
-
-
-description = """
-<div style="border: 3px solid #FF5733; padding: 18px; border-radius: 12px; text-align: center; margin-bottom: 20px;">
-    <h2 style="margin: 0; font-size: 22px; font-weight: bold;">
-        ⚠️ Important Notice
-    </h2>
-    <p style="margin: 10px 0 0 0; font-size: 18px; font-weight: bold;">
-        This is an <u>English</u> AI text detector.<br>
-        The <u>Hungarian</u> AI detector is available at:
-        <a href="https://preds.hu" target="_blank" style="color: #007bff; text-decoration: none;">
-            <b>https://preds.hu</b>
-        </a>
-    </p>
-    <hr style="margin: 15px 0;">
-    <h2 style="margin: 0; font-size: 22px; font-weight: bold;">
-        ⚠️ Fontos információ
-    </h2>
-    <p style="margin: 10px 0 0 0; font-size: 18px; font-weight: bold;">
-        Ez egy <u>angol</u> AI-szövegdetektor.<br>
-        A <u>magyar</u> AI detektor itt érhető el:
-        <a href="https://preds.hu" target="_blank" style="color: #007bff; text-decoration: none;">
-            <b>https://preds.hu</b>
-        </a>
-    </p>
-</div>
-
-
-
-This tool uses the <b>ModernBERT</b> model to identify whether a given text was written by a human or generated by artificial intelligence (AI). It works with a soft voting ensemble using <b>three</b> models, combining their outputs to improve the accuracy.<br>
-<div style="line-height: 1.8;">
-✅ <b>Human Verification:</b> Human-written content is clearly marked.<br>
-🔍 <b>Model Detection:</b> Can identify content from over 40 AI models.<br>
-📈 <b>Accuracy:</b> Works best with longer texts.<br>
-📄 <b>Read more:</b> Our method is detailed in our paper:
-<a href="https://aclanthology.org/2025.genaidetect-1.15/" target="_blank" style="color: #007bff; text-decoration: none;"><b>LINK</b></a>.
-</div>
-<br>
-Paste your text below to analyze its origin.
-"""
-bottom_text = "**Developed by SzegedAI**"
-
-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.",
-"Wines are a fascinating reflection of culture, history, and craftsmanship. They embody a rich diversity shaped by the land, climate, and traditions where they are produced. From the bold reds of Bordeaux to the crisp whites of New Zealand, each bottle tells a unique story. What makes wine so special is its ability to connect people. Whether shared at a family dinner, a celebratory event, or a quiet evening with friends, wine enhances experiences and brings people together. The variety of flavors and aromas, influenced by grape type, fermentation techniques, and aging processes, make wine tasting a complex yet rewarding journey for the senses.",
-"I find artificial intelligence (AI) to be one of the most transformative and fascinating technologies of our time. Its potential spans a wide range of applications, from automating mundane tasks to revolutionizing industries like healthcare, education, and entertainment. AI has already made significant contributions in fields like language processing, image recognition, and decision-making systems, enabling innovations that were once purely science fiction. However, as powerful as AI can be, it also brings challenges and responsibilities. Ethical considerations, such as bias in data, transparency, and the potential for misuse, need to be carefully addressed to ensure fairness and accountability. The rise of generative AI has also sparked debates about creativity, originality, and intellectual property, making it essential to strike a balance between technological advancement and respecting human contributions."
-]
-
-Human_texts = [
-"The present book is intended as a text in basic mathematics. As such, it can have multiple use: for a one-year course in the high schools during the third or fourth year (if possible the third, so that calculus can be taken during the fourth year); for a complementary reference in earlier high school grades (elementary algebra and geometry are covered); for a one-semester course at the college level, to review or to get a firm foundation in the basic mathematics necessary to go ahead in calculus, linear algebra, or other topics. Years ago, the colleges used to give courses in “ college algebra” and other subjects which should have been covered in high school. More recently, such courses have been thought unnecessary, but some experiences I have had show that they are just as necessary as ever. What is happening is that thecolleges are getting a wide variety of students from high schools, ranging from exceedingly well-prepared ones who have had a good first course in calculus, down to very poorly prepared ones.",
-"Fats are rich in energy, build body cells, support brain development of infants, help body processes, and facilitate the absorption and use of fat-soluble vitamins A, D, E, and K. The major component of lipids is glycerol and fatty acids. According to chemical properties, fatty acids can be divided into saturated and unsaturated fatty acids. Generally lipids containing saturated fatty acids are solid at room temperature and include animal fats (butter, lard, tallow, ghee) and tropical oils (palm,coconut, palm kernel). Saturated fats increase the risk of heart disease.",
-"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."]
-
-iface = gr.Blocks(css="""
-    @import url('https://fonts.googleapis.com/css2?family=Roboto+Mono:wght@400;700&display=swap');
-    #text_input_box { border-radius: 10px; border: 2px solid #4CAF50; font-size: 18px; padding: 15px; margin-bottom: 20px; width: 60%; box-sizing: border-box; margin: auto; }
-    .form.svelte-633qhp { background: none; border: none; box-shadow: none; }
-    #result_output_box { border-radius: 10px; border: 2px solid #4CAF50; font-size: 18px; padding: 15px; margin-top: 20px; width: 40%; box-sizing: border-box; text-align: center; margin: auto; }
-    @media (max-width: 768px) { #result_output_box { width: 100%; } #text_input_box{ width: 100%; } }
-    body { font-family: 'Roboto Mono', sans-serif !important; padding: 20px; display: block; justify-content: center; align-items: center; height: 100vh; overflow-y: auto; }
-    .gradio-container { border: 1px solid #4CAF50; border-radius: 15px; padding: 30px; box-shadow: 0px 0px 10px rgba(0,255,0,0.6); max-width: 800px; margin: auto; overflow-y: auto; }
-    h1 { text-align: center; font-size: 32px; font-weight: bold; margin-bottom: 30px; }
-    .highlight-human { color: #4CAF50; font-weight: bold; background: rgba(76, 175, 80, 0.2); padding: 5px; border-radius: 8px; }
-    .highlight-ai { color: #FF5733; font-weight: bold; background: rgba(255, 87, 51, 0.2); padding: 5px; border-radius: 8px; }
-    #bottom_text { text-align: center; margin-top: 50px; font-weight: bold; font-size: 20px; }
-    .block.svelte-11xb1hd{ background: none !important; }
-""")
-
-with iface:
-    gr.Markdown(f"# {title}")
-    gr.Markdown(description)
-    text_input = gr.Textbox(label="", placeholder="Type or paste your content here...", elem_id="text_input_box", lines=5)
-    result_output = gr.Markdown("", elem_id="result_output_box")
-    plot_output = gr.Plot(label="Model Probability Distribution")
-
+# --- 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])
 
-    with gr.Tab("AI text examples"):
-        gr.Examples(AI_texts, inputs=text_input)
-    with gr.Tab("Human text examples"):
-        gr.Examples(Human_texts, inputs=text_input)
-    gr.Markdown(bottom_text, elem_id="bottom_text")
+    gr.Markdown("---")
+    gr.Markdown("**Developed by SzegedAI**")
 
-iface.launch(share=True)
+if __name__ == "__main__":
+    # share=True creates a public URL
+    iface.launch(share=True)