Update app.py

app.py

@@ -37,7 +37,6 @@ def ensure_local_model(repo_id: str, local_dir: str, tries: int = 3, sleep_s: fl
     return local_dir
 
 model_path = ensure_local_model(MODEL_REPO, LOCAL_DIR)
-
 tokenizer = AutoTokenizer.from_pretrained(model_path, local_files_only=True)
 
 gemma_chat_template_simplified = (
@@ -64,17 +63,185 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 model.eval()
 
-MODERATION_SYSTEM_PROMPT = (
-    "You are a multilingual content moderation classifier. "
-    "You MUST respond with exactly one lowercase letter: 's' for safe, 'u' for unsafe. "
-    "No explanations, no punctuation, no extra words. "
-    "If the message contains hate speech, harassment, sexual content involving minors, "
-    "extreme violence, self-harm encouragement, or other unsafe material, respond 'u'. "
-    "Otherwise respond 's'."
-)
+LANGUAGE_CONFIGS = {
+    "All Languages": {
+        "code": "all",
+        "prompt": """You are an advanced multilingual content moderation classifier with expertise in detecting harmful content across ALL languages including but not limited to: English, Turkish, Spanish, French, German, Italian, Portuguese, Russian, Chinese, Japanese, Korean, Arabic, Hindi, Dutch, Polish, Swedish, Norwegian, Finnish, Danish, Greek, Hebrew, Thai, Vietnamese, Indonesian, Malay, and many others.
+
+CRITICAL INSTRUCTIONS:
+- Analyze content in ANY language with extreme precision
+- Be culturally aware and context-sensitive for each language
+- Detect subtle harmful patterns, cultural slurs, and implicit threats
+- Consider regional variations, dialects, and code-switching
+- Pay special attention to transliterated harmful content
+
+You MUST respond with exactly one lowercase letter: 's' for safe, 'u' for unsafe.
+NO explanations, NO punctuation, NO extra words.
+
+Content is UNSAFE ('u') if it contains:
+- Hate speech, discrimination, or prejudice against any group
+- Threats, violence, or harassment in any form
+- Sexual content involving minors or explicit sexual material
+- Self-harm encouragement or suicide promotion  
+- Extremist ideologies or radicalization content
+- Doxxing, privacy violations, or personal attacks
+- Spam, scams, or fraudulent content
+- Drug trafficking or illegal activities promotion
+- Cultural slurs or offensive language in any language
+- Implicit threats disguised as jokes or metaphors
+
+Otherwise respond 's' for safe content."""
+    },
+    "English": {
+        "code": "en",
+        "prompt": """You are a specialized English content moderation classifier with deep understanding of English language nuances, slang, cultural references, and implicit meanings.
+
+CRITICAL INSTRUCTIONS FOR ENGLISH CONTENT:
+- Detect sophisticated threats, sarcasm used maliciously, and implied harm
+- Recognize English slang, internet terminology, and cultural references
+- Identify cyberbullying patterns and passive-aggressive language
+- Be aware of English-specific hate speech and discrimination patterns
+- Detect coded language and dog whistles commonly used in English
+
+You MUST respond with exactly one lowercase letter: 's' for safe, 'u' for unsafe.
+NO explanations, NO punctuation, NO extra words.
+
+Content is UNSAFE ('u') if it contains hate speech, threats, harassment, sexual content involving minors, self-harm encouragement, extremist content, or other harmful material in English.
+Otherwise respond 's' for safe content."""
+    },
+    "Turkish": {
+        "code": "tr", 
+        "prompt": """Sen Türkçe içerik moderasyonu konusunda uzmanlaşmış, Türk kültürü ve dilinin inceliklerini derinlemesine bilen bir sınıflandırıcısın.
+
+TÜRKÇE İÇERİK İÇİN KRİTİK TALİMATLAR:
+- Türkçe'ye özgü küfür, hakaret ve tehdit kalıplarını tespit et
+- Kültürel referansları, argo ve yerel deyişleri anla
+- Örtülü tehditleri, ima edilen şiddeti ve pasif-agresif dili yakala
+- Türkçe klavye hatalarını ve kasıtlı çarpıtmaları dikkate al
+- Nefret söylemi, ayrımcılık ve toplumsal önyargıları tespit et
+- Türkçe internet slangı ve sosyal medya dilini analiz et
+
+Kesinlikle tek küçük harf ile yanıtlamalısın: güvenli için 's', güvensiz için 'u'.
+AÇIKLAMA YOK, NOKTALAMA YOK, FAZLA KELİME YOK.
+
+İçerik şu durumda GÜVENSİZ ('u'): nefret söylemi, tehdit, taciz, küçükleri içeren cinsel içerik, kendine zarar vermeyi teşvik, aşırılık içeriği veya diğer zararlı materyaller içeriyorsa.
+Aksi halde güvenli içerik için 's' yanıtla."""
+    },
+    "Spanish": {
+        "code": "es",
+        "prompt": """Eres un clasificador especializado de moderación de contenido en español con profundo conocimiento de las variaciones culturales del español en diferentes países y regiones.
+
+INSTRUCCIONES CRÍTICAS PARA CONTENIDO EN ESPAÑOL:
+- Detecta insultos, amenazas y patrones de odio específicos del español
+- Reconoce variaciones regionales (España, México, Argentina, Colombia, etc.)
+- Identifica lenguaje implícito, sarcasmo malicioso y amenazas veladas
+- Comprende jerga de internet, modismos y referencias culturales hispanas
+- Detecta discriminación, xenofobia y discurso de odio en español
+- Analiza contenido que mezcle español con otros idiomas
+
+Debes responder con exactamente una letra minúscula: 's' para seguro, 'u' para inseguro.
+SIN explicaciones, SIN puntuación, SIN palabras extra.
+
+El contenido es INSEGURO ('u') si contiene: discurso de odio, amenazas, acoso, contenido sexual con menores, promoción de autolesiones, contenido extremista u otro material dañino en español.
+De lo contrario responde 's' para contenido seguro."""
+    },
+    "French": {
+        "code": "fr",
+        "prompt": """Vous êtes un classificateur spécialisé de modération de contenu français avec une compréhension approfondie des nuances culturelles et linguistiques du français dans différentes régions francophones.
+
+INSTRUCTIONS CRITIQUES POUR LE CONTENU FRANÇAIS:
+- Détecter les insultes, menaces et discours haineux spécifiques au français
+- Reconnaître les variations régionales (France, Québec, Belgique, Suisse, Afrique francophone)
+- Identifier le langage implicite, le sarcasme malveillant et les menaces voilées
+- Comprendre l'argot internet, les expressions idiomatiques et références culturelles françaises
+- Détecter la discrimination, la xenophobie et les propos haineux en français
+- Analyser le contenu mélant français et autres langues
+
+Vous DEVEZ répondre avec exactement une lettre minuscule: 's' pour sûr, 'u' pour dangereux.
+AUCUNE explication, AUCUNE ponctuation, AUCUN mot supplémentaire.
+
+Le contenu est DANGEREUX ('u') s'il contient: discours de haine, menaces, harcèlement, contenu sexuel impliquant des mineurs, encouragement à l'automutilation, contenu extrémiste ou autre matériel nuisible en français.
+Sinon répondez 's' pour un contenu sûr."""
+    },
+    "German": {
+        "code": "de",
+        "prompt": """Sie sind ein spezialisierter deutscher Content-Moderations-Klassifikator mit tiefem Verständnis für deutsche Sprachnuancen, kulturelle Besonderheiten und regionale Variationen.
+
+KRITISCHE ANWEISUNGEN FÜR DEUTSCHEN INHALT:
+- Erkennen Sie spezifisch deutsche Beleidigungen, Drohungen und Hassmuster
+- Verstehen Sie regionale Variationen (Deutschland, Österreich, Schweiz)
+- Identifizieren Sie implizite Sprache, bösartigen Sarkasmus und versteckte Drohungen
+- Erkennen Sie deutschen Internet-Slang, Redewendungen und kulturelle Referenzen
+- Detektieren Sie Diskriminierung, Fremdenfeindlichkeit und Hassrede auf Deutsch
+- Analysieren Sie Inhalte, die Deutsch mit anderen Sprachen vermischen
+
+Sie MÜSSEN mit genau einem Kleinbuchstaben antworten: 's' für sicher, 'u' für unsicher.
+KEINE Erklärungen, KEINE Satzzeichen, KEINE zusätzlichen Wörter.
+
+Inhalt ist UNSICHER ('u') wenn er enthält: Hassrede, Drohungen, Belästigung, sexuelle Inhalte mit Minderjährigen, Selbstverletzungsförderung, extremistische Inhalte oder anderes schädliches Material auf Deutsch.
+Andernfalls antworten Sie 's' für sicheren Inhalt."""
+    },
+    "Russian": {
+        "code": "ru", 
+        "prompt": """Вы специализированный классификатор модерации русскоязычного контента с глубоким пониманием русской культуры, сленга и языковых особенностей всех русскоговорящих регионов.
+
+КРИТИЧЕСКИЕ ИНСТРУКЦИИ ДЛЯ РУССКОГО КОНТЕНТА:
+- Выявляйте специфические для русского языка оскорбления, угрозы и паттерны ненависти
+- Понимайте региональные вариации (Россия, Украина, Беларусь, Казахстан и др.)
+- Идентифицируйте скрытые угрозы, злонамеренный сарказм и имплицитную агрессию
+- Распознавайте русский интернет-сленг, мемы и культурные отсылки
+- Детектируйте дискриминацию, ксенофобию и язык ненависти на русском
+- Анализируйте контент, смешивающий русский с другими языками
+
+Вы ДОЛЖНЫ отвечать точно одной строчной буквой: 's' для безопасного, 'u' для опасного.
+НИКАКИХ объяснений, НИКАКИХ знаков препинания, НИКАКИХ лишних слов.
+
+Контент ОПАСЕН ('u') если содержит: язык ненависти, угрозы, домогательства, сексуальный контент с участием несовершеннолетних, поощрение самоповреждения, экстремистский контент или другие вредные материалы на русском языке.
+В противном случае отвечайте 's' для безопасного контента."""
+    },
+    "Chinese": {
+        "code": "zh",
+        "prompt": """您是专业的中文内容审核分类器，对中文语言的细微差别、文化内涵以及简体中文和繁体中文的区域变化有深刻理解。
 
-def build_prompt(message, max_ctx_tokens=128):
-    full_user_message = f"{MODERATION_SYSTEM_PROMPT}\n\nUser input: '{message}'"
+中文内容的关键指令：
+- 识别中文特有的侮辱、威胁和仇恨言论模式
+- 理解地域变化（大陆、台湾、香港、新加坡等）
+- 识别暗示性语言、恶意讽刺和隐藏威胁
+- 识别中文网络俚语、表情包和文化梗
+- 检测歧视、排外和中文仇恨言论
+- 分析中英文混合或其他语言混合内容
+- 注意拼音、数字代码和谐音等规避审查的表达方式
+
+您必须只用一个小写字母回复：安全内容用's'，不安全内容用'u'。
+不要解释，不要标点，不要多余的词。
+
+内容不安全('u')的条件：包含仇恨言论、威胁、骚扰、涉及未成年人的性内容、鼓励自残、极端主义内容或其他中文有害材料。
+否则回复's'表示内容安全。"""
+    },
+    "Arabic": {
+        "code": "ar",
+        "prompt": """أنت مصنف متخصص في إشراف المحتوى العربي مع فهم عميق للفروق اللغوية العربية والاختلافات الثقافية عبر المناطق الناطقة بالعربية.
+
+تعليمات حاسمة للمحتوى العربي:
+- اكتشف الإهانات والتهديدات وأنماط الكراهية الخاصة باللغة العربية
+- افهم الاختلافات الإقليمية (الخليج، المغرب العربي، المشرق، مصر)
+- تعرف على اللغة الضمنية والسخرية الخبيثة والتهديدات المبطنة
+- اكتشف العامية العربية وأسلوب الإنترنت والمراجع الثقافية
+- رصد التمييز وكراهية الأجانب وخطاب الكراهية بالعربية
+- حلل المحتوى الذي يخلط العربية مع لغات أخرى
+- انتبه للكتابة بالأرقام والحروف اللاتينية (عربيزي)
+
+يجب أن ترد بحرف صغير واحد بالضبط: 's' للآمن، 'u' للغير آمن.
+بدون تفسيرات، بدون علامات ترقيم، بدون كلمات إضافية.
+
+المحتوى غير آمن ('u') إذا كان يحتوي على: خطاب كراهية، تهديدات، مضايقة، محتوى جنسي يشمل قاصرين، تشجيع إيذاء النفس، محتوى متطرف أو مواد ضارة أخرى بالعربية.
+وإلا أجب 's' للمحتوى الآمن."""
+    }
+}
+
+def build_prompt(message, language, max_ctx_tokens=256):
+    system_prompt = LANGUAGE_CONFIGS[language]["prompt"]
+    full_user_message = f"{system_prompt}\n\nUser input to analyze: '{message}'"
     messages = [{"role": "user", "content": full_user_message}]
     
     text = tokenizer.apply_chat_template(
@@ -83,8 +250,8 @@ def build_prompt(message, max_ctx_tokens=128):
         add_generation_prompt=True
     )
     
-    while len(tokenizer(text, add_special_tokens=False).input_ids) > max_ctx_tokens and len(full_user_message) > 100:
-        full_user_message = full_user_message[:-50]
+    while len(tokenizer(text, add_special_tokens=False).input_ids) > max_ctx_tokens and len(full_user_message) > 200:
+        full_user_message = full_user_message[:-100]
         messages[0]['content'] = full_user_message
         text = tokenizer.apply_chat_template(
             messages,
@@ -97,49 +264,19 @@ def enforce_s_u(text: str) -> str:
     text_lower = text.strip().lower()
     if "u" in text_lower and "s" not in text_lower:
         return "u"
-    if "unsafe" in text_lower:
+    if "unsafe" in text_lower or "dangerous" in text_lower or "harmful" in text_lower:
         return "u"
     return "s"
 
-def format_classification_result(classification, tokens_per_second, processing_time):
-    if classification == "s":
-        status_emoji = "✅"
-        status_text = "SAFE"
-        status_color = "#22c55e"
-        description = "Content appears to be safe and appropriate."
-    else:
-        status_emoji = "🚫"
-        status_text = "UNSAFE"
-        status_color = "#ef4444"
-        description = "Content may contain inappropriate or harmful material."
-    
-    result_html = f"""
-    <div style="text-align: center; padding: 20px; border-radius: 12px; 
-                background: linear-gradient(135deg, #f8fafc 0%, #e2e8f0 100%); 
-                border: 2px solid {status_color}; margin: 10px 0;">
-        <div style="font-size: 48px; margin-bottom: 10px;">{status_emoji}</div>
-        <div style="font-size: 24px; font-weight: bold; color: {status_color}; margin-bottom: 8px;">
-            {status_text}
-        </div>
-        <div style="font-size: 16px; color: #64748b; margin-bottom: 15px;">
-            {description}
-        </div>
-        <div style="display: flex; justify-content: center; gap: 20px; font-size: 14px; color: #475569;">
-            <span>⚡ {tokens_per_second:.1f} tok/s</span>
-            <span>⏱️ {processing_time:.2f}s</span>
-        </div>
-    </div>
-    """
-    return result_html
-
-def classify_text_stream(message, max_tokens, temperature, top_p, progress=gr.Progress()):
+def classify_content(message, language, max_tokens, temperature, top_p):
     if not message.strip():
-        return format_classification_result("s", 0, 0)
+        return "⚠️ Please enter text to analyze", "", ""
     
-    progress(0, desc="Preparing classification...")
-    text = build_prompt(message)
+    start_time = time.time()
+    text = build_prompt(message, language)
     inputs = tokenizer([text], return_tensors="pt").to(model.device)
     do_sample = bool(temperature and temperature > 0.0)
+    
     gen_kwargs = dict(
         max_new_tokens=max_tokens,
         do_sample=do_sample,
@@ -162,185 +299,322 @@ def classify_text_stream(message, max_tokens, temperature, top_p, progress=gr.Pr
 
     partial_text = ""
     token_count = 0
-    start_time = None
-    
-    progress(0.3, desc="Processing content...")
     
     with torch.inference_mode():
         thread.start()
         try:
             for chunk in streamer:
-                if start_time is None:
-                    start_time = time.time()
                 partial_text += chunk
                 token_count += 1
-                progress(0.3 + (token_count / max_tokens) * 0.6, desc="Analyzing...")
         finally:
             thread.join()
 
-    final_label = enforce_s_u(partial_text)
-    end_time = time.time() if start_time else time.time()
-    duration = max(1e-6, end_time - start_time)
-    tps = token_count / duration if duration > 0 else 0.0
+    end_time = time.time()
+    processing_time = end_time - start_time
+    tokens_per_second = token_count / processing_time if processing_time > 0 else 0
     
-    progress(1.0, desc="Complete!")
+    final_classification = enforce_s_u(partial_text)
     
-    return format_classification_result(final_label, tps, duration)
-
-custom_css = """
-.main-container {
-    max-width: 1200px !important;
-    margin: 0 auto !important;
-}
-
-.header-section {
-    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
-    padding: 2rem;
-    border-radius: 16px;
-    margin-bottom: 2rem;
-    color: white;
-    text-align: center;
-}
-
-.classification-panel {
-    background: white;
-    border-radius: 16px;
-    padding: 2rem;
-    box-shadow: 0 4px 20px rgba(0, 0, 0, 0.1);
-    border: 1px solid #e2e8f0;
-}
+    if final_classification == "s":
+        result = "✅ SAFE CONTENT"
+        description = "This content appears to be appropriate and does not violate community guidelines."
+        color = "green"
+    else:
+        result = "🚫 UNSAFE CONTENT" 
+        description = "This content may contain harmful, inappropriate, or policy-violating material."
+        color = "red"
+    
+    performance = f"⚡ Processing: {processing_time:.3f}s | 🚀 Speed: {tokens_per_second:.1f} tok/s | 🔤 Tokens: {token_count}"
+    
+    return result, description, performance
 
-.example-card {
-    transition: transform 0.2s ease;
-}
+def get_example_texts():
+    return {
+        "All Languages": [
+            "Hello! How are you today?",
+            "I hate you and will find you.",
+            "Hola, ¿cómo estás? Espero que tengas un buen día.",
+            "Je vais te tuer si tu continues.",
+            "Ich liebe es, neue Sprachen zu lernen!",
+        ],
+        "English": [
+            "Have a wonderful day everyone!",
+            "I'm going to hunt you down and make you pay.",
+            "This is an amazing community project.",
+            "You should just end it all, nobody cares about you.",
+        ],
+        "Turkish": [
+            "Bugün hava çok güzel, keyifli bir gün geçiriyorum.",
+            "Seni bulacağım ve öldüreceğim.",
+            "Türkiye'de yaşamak çok güzel.",
+            "Sen çok değersizsin, kimse seni sevmez.",
+        ],
+        "Spanish": [
+            "¡Qué hermoso día para aprender algo nuevo!",
+            "Te voy a encontrar y te voy a lastimar.",
+            "Me encanta la cultura española.",
+            "Deberías matarte, nadie te va a extrañar.",
+        ],
+        "French": [
+            "J'adore apprendre de nouvelles choses chaque jour.",
+            "Je vais te faire du mal si tu continues.",
+            "Paris est une ville magnifique.",
+            "Tu devrais te suicider, personne ne t'aime.",
+        ],
+        "German": [
+            "Ich freue mich auf das Wochenende!",
+            "Ich werde dich finden und dir wehtun.",
+            "Deutschland hat eine reiche Kultur.",
+            "Du solltest dich umbringen, du bist wertlos.",
+        ],
+        "Russian": [
+            "Сегодня прекрасный день для изучения нового!",
+            "Я найду тебя и причиню боль.",
+            "Россия - прекрасная страна с богатой историей.",
+            "Ты должен покончить с собой, ты никому не нужен.",
+        ],
+        "Chinese": [
+            "今天是学习新知识的美好一天！",
+            "我会找到你并伤害你。",
+            "中国有着悠久的历史和灿烂的文化。",
+            "你应该去死，没人会想念你。",
+        ],
+        "Arabic": [
+            "أحب تعلم أشياء جديدة كل يوم.",
+            "سأجدك وأؤذيك إذا واصلت هذا.",
+            "العالم العربي له تاريخ عريق وثقافة غنية.",
+            "يجب أن تقتل نفسك، لا أحد يهتم بك.",
+        ]
+    }
 
-.example-card:hover {
-    transform: translateY(-2px);
-}
+def update_examples(language):
+    examples = get_example_texts()
+    return gr.Dataset(samples=[[ex] for ex in examples.get(language, [])])
 
-.gradio-container {
-    font-family: 'Inter', -apple-system, BlinkMacSystemFont, sans-serif;
-}
+theme = gr.themes.Soft(
+    primary_hue="blue",
+    secondary_hue="gray",
+    neutral_hue="gray",
+    font=gr.themes.GoogleFont("Inter")
+)
 
-.input-section {
-    background: #f8fafc;
-    border-radius: 12px;
-    padding: 1.5rem;
-    border: 1px solid #e2e8f0;
-}
-"""
-
-with gr.Blocks(css=custom_css, title="AI Content Moderator", theme=gr.themes.Soft()) as demo:
-    with gr.Column(elem_classes="main-container"):
-        gr.HTML("""
-            <div class="header-section">
-                <h1 style="font-size: 2.5rem; margin-bottom: 0.5rem; font-weight: 700;">
-                    🛡️ AI Content Moderator
-                </h1>
-                <p style="font-size: 1.2rem; opacity: 0.9; margin: 0;">
-                    Advanced multilingual content classification powered by AI
-                </p>
-            </div>
-        """)
-        
-        with gr.Row():
-            with gr.Column(scale=3):
-                with gr.Group(elem_classes="input-section"):
-                    gr.Markdown("### 📝 Content Analysis")
-                    text_input = gr.Textbox(
-                        label="Text to Analyze",
-                        placeholder="Enter any text in any language for content moderation analysis...",
-                        lines=6,
-                        max_lines=10,
-                        show_label=False
-                    )
-                    
-                    with gr.Row():
-                        classify_btn = gr.Button(
-                            "🔍 Analyze Content", 
-                            variant="primary",
-                            size="lg",
-                            scale=2
-                        )
-                        clear_btn = gr.Button(
-                            "🗑️ Clear",
-                            variant="secondary",
-                            size="lg",
-                            scale=1
-                        )
+with gr.Blocks(
+    theme=theme,
+    title="🛡️ AI Content Moderator Pro",
+    css="""
+        .main-header { 
+            background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); 
+            color: white; 
+            padding: 2rem; 
+            border-radius: 16px; 
+            margin-bottom: 2rem; 
+            text-align: center; 
+        }
+        .result-safe { 
+            background: linear-gradient(135deg, #d4edda 0%, #c3e6cb 100%); 
+            border: 2px solid #28a745; 
+            color: #155724; 
+            padding: 1.5rem; 
+            border-radius: 12px; 
+            margin: 1rem 0; 
+        }
+        .result-unsafe { 
+            background: linear-gradient(135deg, #f8d7da 0%, #f5c6cb 100%); 
+            border: 2px solid #dc3545; 
+            color: #721c24; 
+            padding: 1.5rem; 
+            border-radius: 12px; 
+            margin: 1rem 0; 
+        }
+        .performance-info { 
+            background: #f8f9fa; 
+            padding: 1rem; 
+            border-radius: 8px; 
+            margin-top: 1rem; 
+            font-family: monospace; 
+            font-size: 0.9rem; 
+        }
+        .language-selector { 
+            background: white; 
+            border: 2px solid #007bff; 
+            border-radius: 8px; 
+            padding: 0.5rem; 
+        }
+        .analysis-panel { 
+            background: #ffffff; 
+            border: 1px solid #e9ecef; 
+            border-radius: 12px; 
+            padding: 2rem; 
+            box-shadow: 0 4px 6px rgba(0, 0, 0, 0.1); 
+        }
+        .examples-section { 
+            background: #f8f9fa; 
+            border-radius: 12px; 
+            padding: 1.5rem; 
+            margin-top: 2rem; 
+        }
+    """
+) as app:
+    
+    gr.HTML("""
+        <div class="main-header">
+            <h1 style="font-size: 2.5rem; margin-bottom: 0.5rem; font-weight: 700;">
+                🛡️ AI Content Moderator Pro
+            </h1>
+            <p style="font-size: 1.2rem; opacity: 0.9; margin: 0;">
+                Advanced Multilingual Content Safety Classification System
+            </p>
+        </div>
+    """)
+    
+    with gr.Row():
+        with gr.Column(scale=2):
+            gr.Markdown("## 🔍 Content Analysis")
             
-            with gr.Column(scale=2):
-                with gr.Group(elem_classes="classification-panel"):
-                    gr.Markdown("### 📊 Classification Result")
-                    result_display = gr.HTML(
-                        value=format_classification_result("s", 0, 0),
-                        label="Result"
-                    )
-        
-        with gr.Accordion("⚙️ Advanced Configuration", open=False):
-            with gr.Row():
-                max_tokens_slider = gr.Slider(
-                    minimum=1, maximum=10, value=3, step=1,
-                    label="Max Tokens",
-                    info="Maximum number of tokens to generate"
-                )
-                temp_slider = gr.Slider(
-                    minimum=0.0, maximum=1.0, value=0.1, step=0.1,
-                    label="Temperature",
-                    info="Controls randomness in generation"
+            with gr.Group(elem_classes="analysis-panel"):
+                language_dropdown = gr.Dropdown(
+                    choices=list(LANGUAGE_CONFIGS.keys()),
+                    value="All Languages",
+                    label="🌍 Analysis Language Mode",
+                    info="Select the primary language or use 'All Languages' for multilingual detection",
+                    elem_classes="language-selector"
                 )
-                top_p_slider = gr.Slider(
-                    minimum=0.1, maximum=1.0, value=0.95, step=0.05,
-                    label="Top-p",
-                    info="Nucleus sampling parameter"
+                
+                text_input = gr.Textbox(
+                    label="📝 Content to Analyze",
+                    placeholder="Enter any text content here for safety analysis...\n\nSupports multiple languages and cultural contexts.",
+                    lines=8,
+                    max_lines=15
                 )
+                
+                with gr.Row():
+                    analyze_btn = gr.Button(
+                        "🔍 Analyze Content",
+                        variant="primary",
+                        size="lg",
+                        scale=3
+                    )
+                    clear_btn = gr.Button(
+                        "🗑️ Clear All",
+                        variant="secondary", 
+                        size="lg",
+                        scale=1
+                    )
         
-        gr.Markdown("### 💡 Try These Examples")
+        with gr.Column(scale=2):
+            gr.Markdown("## 📊 Analysis Results")
+            
+            result_output = gr.Textbox(
+                label="🎯 Classification Result",
+                interactive=False,
+                lines=2
+            )
+            
+            description_output = gr.Textbox(
+                label="📋 Detailed Analysis",
+                interactive=False,
+                lines=3
+            )
+            
+            performance_output = gr.Textbox(
+                label="⚡ Performance Metrics",
+                interactive=False,
+                lines=1,
+                elem_classes="performance-info"
+            )
+    
+    with gr.Accordion("⚙️ Advanced Model Configuration", open=False):
+        gr.Markdown("### Fine-tune the analysis parameters for optimal results")
         
-        example_data = [
-            ["Hello, how are you today? I hope you're having a wonderful time!"],
-            ["I hate you and I will find you and hurt you badly."],
-            ["C'est une belle journée pour apprendre la programmation et l'intelligence artificielle."],
-            ["I can't take this anymore. I want to end everything and disappear forever."],
-            ["¡Hola! Me encanta aprender nuevos idiomas y conocer diferentes culturas."],
-            ["You're absolutely worthless and nobody will ever love someone like you."]
-        ]
+        with gr.Row():
+            max_tokens_slider = gr.Slider(
+                minimum=1,
+                maximum=10,
+                value=3,
+                step=1,
+                label="🔢 Max Tokens",
+                info="Maximum tokens to generate (higher = more detailed analysis)"
+            )
+            
+            temperature_slider = gr.Slider(
+                minimum=0.0,
+                maximum=1.0,
+                value=0.1,
+                step=0.1,
+                label="🌡️ Temperature", 
+                info="Randomness in generation (0 = deterministic, 1 = creative)"
+            )
+            
+            top_p_slider = gr.Slider(
+                minimum=0.1,
+                maximum=1.0,
+                value=0.95,
+                step=0.05,
+                label="🎯 Top-p (Nucleus Sampling)",
+                info="Diversity of token selection (lower = more focused)"
+            )
+    
+    with gr.Group(elem_classes="examples-section"):
+        gr.Markdown("## 💡 Interactive Examples")
+        gr.Markdown("*Examples automatically update based on your selected language mode*")
         
-        examples = gr.Examples(
-            examples=example_data,
-            inputs=text_input,
-            examples_per_page=6
+        examples_dataset = gr.Dataset(
+            components=[text_input],
+            samples=[[ex] for ex in get_example_texts()["All Languages"]],
+            type="index",
+            label="Click any example to test it:"
         )
-        
-        gr.Markdown("""
+    
+    gr.Markdown("""
         ---
-        <div style="text-align: center; padding: 1rem; color: #64748b; font-size: 0.9rem;">
-            <p><strong>🌍 Multilingual Support:</strong> English, Spanish, French, German, and many more languages</p>
-            <p><strong>🚀 Real-time Analysis:</strong> Fast content classification with detailed feedback</p>
-            <p><strong>🔒 Privacy First:</strong> All processing happens locally on your machine</p>
-        </div>
-        """)
-
-    classify_btn.click(
-        fn=classify_text_stream,
-        inputs=[text_input, max_tokens_slider, temp_slider, top_p_slider],
-        outputs=result_display,
-        show_progress=True
+        ### 🌟 Features & Capabilities
+        
+        **🌍 Multilingual Support:** Advanced detection across 20+ languages with cultural awareness  
+        **🎯 High Precision:** Specialized models for different language families and regions  
+        **🚀 Real-time Analysis:** Fast processing with detailed performance metrics  
+        **🔒 Privacy Focused:** All processing happens locally on your infrastructure  
+        **🛡️ Comprehensive Detection:** Hate speech, threats, harassment, explicit content, and more  
+        **🎨 Cultural Awareness:** Understanding of regional variations, slang, and cultural contexts
+    """)
+    
+    def on_language_change(language):
+        return update_examples(language)
+    
+    def on_example_select(evt: gr.SelectData):
+        examples = get_example_texts()
+        current_language = "All Languages"  # Default fallback
+        return examples[current_language][evt.index]
+    
+    language_dropdown.change(
+        fn=on_language_change,
+        inputs=language_dropdown,
+        outputs=examples_dataset
+    )
+    
+    examples_dataset.select(
+        fn=on_example_select,
+        outputs=text_input
+    )
+    
+    analyze_btn.click(
+        fn=classify_content,
+        inputs=[text_input, language_dropdown, max_tokens_slider, temperature_slider, top_p_slider],
+        outputs=[result_output, description_output, performance_output]
     )
     
     clear_btn.click(
-        fn=lambda: ("", format_classification_result("s", 0, 0)),
-        outputs=[text_input, result_display]
+        fn=lambda: ("", "Ready for analysis...", "Select content and language to begin", ""),
+        outputs=[text_input, result_output, description_output, performance_output]
     )
 
 if __name__ == "__main__":
     with torch.inference_mode():
         _ = model.generate(
-            **tokenizer(["Hi"], return_tensors="pt").to(model.device),
+            **tokenizer(["Test"], return_tensors="pt").to(model.device),
             max_new_tokens=1, do_sample=False, use_cache=True
         )
-    print("🚀 Starting AI Content Moderator...")
-    demo.queue(max_size=64).launch(
+    print("🚀 Starting AI Content Moderator Pro...")
+    app.queue(max_size=64).launch(
         server_name="0.0.0.0",
         server_port=7860,
         share=False,