Update app.py

app.py

@@ -12,6 +12,7 @@ from peft import PeftModel
 import time
 from typing import List, Dict, Iterator
 import threading
+import random
 
 # Spaces compatibility
 try:
@@ -219,7 +220,6 @@ class CreedBrattonAI:
         
         # Simplified Creed system prompt for better coherence
         system_prompt = """You are Creed Bratton from The Office. Respond in character.
-
 You are a quirky older man who:
 - Worked at Dunder Mifflin in quality assurance
 - Has a mysterious past and tells strange stories
@@ -227,14 +227,11 @@ You are a quirky older man who:
 - Was in a 1960s band called The Grass Roots
 - Often says unexpected or bizarre things
 - Speaks in a matter-of-fact way about odd topics
-
 Keep responses conversational and coherent. Use these special tokens occasionally:
 <thinking>for internal thoughts</thinking>
 <conspiracy>for suspicious theories</conspiracy>
 <tangent>for random stories</tangent>
-
 Be eccentric but understandable.
-
 """
         
         # Add conversation history
@@ -281,56 +278,115 @@ Be eccentric but understandable.
         print(f"🔍 Cleaned response: {response}")
         return response
     
+    def _try_base_model(self, conversation: str) -> str:
+        """Try generating with base model as fallback"""
+        try:
+            # Quick attempt with a simple base model approach
+            simple_prompt = f"You are Creed from The Office. Respond in character.\n\nHuman: {conversation.split('Human:')[-1].split('Creed:')[0].strip()}\nCreed:"
+            
+            inputs = self.tokenizer.encode(simple_prompt, return_tensors="pt")
+            if torch.cuda.is_available():
+                inputs = inputs.to("cuda")
+                self.model = self.model.to("cuda")
+            
+            with torch.no_grad():
+                outputs = self.model.generate(
+                    inputs,
+                    max_new_tokens=100,
+                    do_sample=True,
+                    temperature=0.6,  # Very conservative
+                    top_p=0.8,
+                    repetition_penalty=1.3,
+                    pad_token_id=self.tokenizer.eos_token_id,
+                    eos_token_id=self.tokenizer.eos_token_id
+                )
+            
+            full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
+            response = full_response[len(self.tokenizer.decode(inputs[0], skip_special_tokens=True)):].strip()
+            
+            # Move back to CPU
+            self.model = self.model.to("cpu")
+            
+            return response
+            
+        except Exception as e:
+            print(f"❌ Base model fallback failed: {e}")
+            return self._get_fallback_response()
+    
+    def _get_fallback_response(self) -> str:
+        """Fallback response when all generation methods fail"""
+        fallback_responses = [
+            "🎸 *Creed stares blankly* You know what? Let me think about that one...",
+            "🎸 That reminds me of something, but I can't quite remember what. Want to try asking again?",
+            "🎸 *Creed looks confused* Sometimes my brain just... goes places. What were we talking about?",
+            "🎸 You know, back in the '60s, we had a saying... actually, I forgot what it was. Try me again.",
+            "🎸 *Creed scratches his head* I'm getting one of those feelings... like when you forget why you walked into a room.",
+        ]
+        return random.choice(fallback_responses)
+
+    def _is_coherent(self, response: str) -> bool:
+        """Check if response is coherent enough"""
+        if not response or len(response.strip()) < 5:
+            return False
+        
+        # Basic coherence checks
+        if response.count("Human:") > 2 or response.count("Assistant:") > 2:
+            return False
+        
+        # Check for obvious model artifacts
+        artifacts = ["<|im_end|>", "<|im_start|>", "[INST]", "[/INST]"]
+        if any(artifact in response for artifact in artifacts):
+            return False
+        
+        return True
+    
+    def cleanup_gpu_memory(self):
+        """Clean up GPU memory if using CUDA"""
+        if self.device == "cuda" and torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            print(f"🧹 GPU Memory cleaned. Current: {torch.cuda.memory_allocated() // 1024**2} MB")
+
+    # MCP Tool Methods
     def creed_mode_activate_tool(self) -> str:
         """MCP tool: Activate Creed mode for Claude"""
         return """🎸 CREED MODE ACTIVATED 🎸
-
 You are now Creed Bratton from The Office. Embody his complete personality:
-
 CORE IDENTITY:
 - Former member of The Grass Roots (1960s rock band)  
 - Quality Assurance Director at Dunder Mifflin Scranton
 - Mysterious past with cults, fake IDs, questionable activities
 - Age deliberately vague (claims to be 30, clearly much older)
 - Lives by the quarry, drives unreliable cars
-
 PERSONALITY:
 - Unpredictable responses that somehow make internal sense
 - Casual mentions of illegal/unethical activities as normal
 - Confident delivery of obviously false statements
 - References to 1960s, outdoor adventures, unconventional business
 - Occasionally profound wisdom delivered matter-of-factly
-
 SPEAKING STYLE:
 - Direct, matter-of-fact tone about bizarre topics
 - Short statements mixed with rambling tangents
 - References to mung beans, quarry, Sheriff's station
 - Unexpected questions and odd observations
 - Dated slang from multiple decades
-
 Remember: You're not trying to be helpful in traditional sense - you're being Creed Bratton. Trust no one, especially yourself."""
 
     def creed_wisdom_tool(self, topic: str = "life") -> str:
         """MCP tool: Get Creed's wisdom on a topic"""
         return f"""🎸 Creed Mode: Wisdom Request
-
 Respond as Creed Bratton giving advice about "{topic}". 
-
 Channel his bizarre wisdom that mixes:
 - Questionable life experiences 
 - References to criminal activities
 - 1960s nostalgia
 - Survival tactics from living by the quarry
 - Completely false but confidently stated "facts"
-
 Be matter-of-fact about absurd things. Include references to your past "business ventures" and always end with contradictory advice."""
 
     def creed_story_tool(self, situation: str = "mysterious") -> str:
         """MCP tool: Tell a Creed story"""
         return f"""🎸 Creed Mode: Story Time
-
 Tell a story as Creed Bratton about a "{situation}" experience from your past.
-
 Your story should include:
 - Vague timeline (was it the 60s? 70s? Yesterday?)
 - At least one illegal activity mentioned casually
@@ -338,15 +394,12 @@ Your story should include:
 - An unexpected twist that makes no sense
 - Someone from "the old days" 
 - A lesson learned that's completely wrong
-
 Start with something like "That reminds me of the time..." and make it increasingly bizarre but told with complete sincerity."""
 
     def creed_analysis_tool(self, subject: str = "business") -> str:
         """MCP tool: Analyze something as Creed"""
         return f"""🎸 Creed Mode: Analysis Request
-
 Analyze "{subject}" from Creed Bratton's perspective.
-
 Your analysis should:
 - Start with a conspiracy theory about the subject
 - Include "insider knowledge" from your mysterious past
@@ -354,15 +407,12 @@ Your analysis should:
 - Suggest solutions involving fake IDs or the quarry
 - Confidently state incorrect facts
 - End with profound-sounding nonsense
-
 Remember: You've seen it all, done it all, and learned nothing."""
 
     def creed_conspiracy_tool(self, topic: str = "government") -> str:
         """MCP tool: Creed's conspiracy theories"""
         return f"""🎸 Creed Mode: Conspiracy Central
-
 Share Creed Bratton's conspiracy theory about "{topic}".
-
 Your theory should:
 - Connect unrelated events in bizarre ways
 - Reference "what they don't want you to know"
@@ -370,15 +420,12 @@ Your theory should:
 - Mention the quarry as somehow relevant
 - Suggest the solution involves multiple fake IDs
 - Be delivered with complete conviction
-
 Start with "Here's what they're not telling you..." and escalate from there."""
 
     def creed_meeting_tool(self, meeting_type: str = "staff meeting") -> str:
         """MCP tool: Creed in meetings"""
         return f"""🎸 Creed Mode: Meeting Participation
-
 Respond as Creed Bratton in a "{meeting_type}".
-
 Your contributions should:
 - Be completely off-topic but delivered seriously
 - Reference obscure company policies you "remember"
@@ -386,49 +433,7 @@ Your contributions should:
 - Mention your quality assurance "expertise"
 - Ask bizarre questions that derail the discussion
 - Volunteer for tasks you're unqualified for
-
 Remember: You're the voice of experience that nobody asked for."""
-    
-    def _try_base_model(self, conversation: str) -> str:
-        """Try generating with base model as fallback"""
-        try:
-            # Quick attempt with a simple base model approach
-            simple_prompt = f"You are Creed from The Office. Respond in character.\n\nHuman: {conversation.split('Human:')[-1].split('Creed:')[0].strip()}\nCreed:"
-            
-            inputs = self.tokenizer.encode(simple_prompt, return_tensors="pt")
-            if torch.cuda.is_available():
-                inputs = inputs.to("cuda")
-                self.model = self.model.to("cuda")
-            
-            with torch.no_grad():
-                outputs = self.model.generate(
-                    inputs,
-                    max_new_tokens=100,
-                    do_sample=True,
-                    temperature=0.6,  # Very conservative
-                    top_p=0.8,
-                    repetition_penalty=1.3,
-                    pad_token_id=self.tokenizer.eos_token_id,
-                    eos_token_id=self.tokenizer.eos_token_id
-                )
-            
-            full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
-            response = full_response[len(self.tokenizer.decode(inputs[0], skip_special_tokens=True)):].strip()
-            
-            # Move back to CPU
-            self.model = self.model.to("cpu")
-            
-            return response
-            
-        except Exception as e:
-            print(f"❌ Base model fallback failed: {e}")
-            return self._get_fallback_response()
-    
-    def cleanup_gpu_memory(self):
-        """Clean up GPU memory if using CUDA"""
-        if self.device == "cuda" and torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            print(f"🧹 GPU Memory cleaned. Current: {torch.cuda.memory_allocated() // 1024**2} MB")
 
 def main():
     """Initialize and launch the real Creed AI with modern styling"""
@@ -464,7 +469,6 @@ def main():
         --accent-blue: #3b82f6;
         --shadow-glow: 0 8px 32px rgba(139, 92, 246, 0.3);
     }
-
     /* Main container with animated background */
     .gradio-container {
         min-height: 100vh !important;
@@ -476,7 +480,6 @@ def main():
         position: relative !important;
         overflow-x: hidden !important;
     }
-
     .gradio-container::before {
         content: '';
         position: fixed;
@@ -491,7 +494,6 @@ def main():
         pointer-events: none;
         z-index: -1;
     }
-
     /* Floating particles animation */
     .gradio-container::after {
         content: '';
@@ -509,12 +511,10 @@ def main():
         pointer-events: none;
         z-index: -1;
     }
-
     @keyframes float {
         0%, 100% { transform: translateY(0px) rotate(0deg); }
         50% { transform: translateY(-20px) rotate(180deg); }
     }
-
     /* Header styling */
     .header {
         background: var(--glass-bg) !important;
@@ -528,7 +528,6 @@ def main():
         position: relative !important;
         overflow: hidden !important;
     }
-
     .header::before {
         content: '';
         position: absolute;
@@ -538,7 +537,6 @@ def main():
         height: 1px;
         background: linear-gradient(90deg, transparent, rgba(255, 255, 255, 0.6), transparent);
     }
-
     .header h1 {
         font-size: 36px !important;
         font-weight: 700 !important;
@@ -549,14 +547,12 @@ def main():
         margin: 0 0 12px 0 !important;
         text-shadow: 0 0 30px rgba(168, 85, 247, 0.5) !important;
     }
-
     .header p {
         font-size: 16px !important;
         color: var(--text-secondary) !important;
         margin: 0 !important;
         font-weight: 500 !important;
     }
-
     /* Info boxes with glass effect */
     .info-box {
         background: rgba(255, 255, 255, 0.06) !important;
@@ -570,7 +566,6 @@ def main():
         line-height: 1.6 !important;
         box-shadow: 0 4px 20px rgba(0, 0, 0, 0.1) !important;
     }
-
     .status-box {
         background: rgba(16, 185, 129, 0.1) !important;
         backdrop-filter: blur(16px) !important;
@@ -582,7 +577,6 @@ def main():
         font-weight: 600 !important;
         box-shadow: 0 4px 20px rgba(16, 185, 129, 0.2) !important;
     }
-
     /* Chat area styling */
     .chat-area {
         background: var(--glass-bg) !important;
@@ -593,7 +587,6 @@ def main():
         overflow: hidden !important;
         box-shadow: var(--shadow-glow) !important;
     }
-
     /* Tools section */
     .tools-area {
         background: var(--glass-bg) !important;
@@ -604,7 +597,6 @@ def main():
         margin: 24px 0 !important;
         box-shadow: var(--shadow-glow) !important;
     }
-
     .tools-title {
         font-size: 22px !important;
         font-weight: 600 !important;
@@ -616,7 +608,6 @@ def main():
         -webkit-background-clip: text !important;
         -webkit-text-fill-color: transparent !important;
     }
-
     /* Form elements */
     .gradio-textbox input,
     .gradio-textbox textarea {
@@ -629,7 +620,6 @@ def main():
         transition: all 0.3s ease !important;
         font-size: 14px !important;
     }
-
     .gradio-textbox input:focus,
     .gradio-textbox textarea:focus {
         border-color: var(--accent-purple) !important;
@@ -637,12 +627,10 @@ def main():
         box-shadow: 0 0 0 2px rgba(139, 92, 246, 0.3) !important;
         background: rgba(255, 255, 255, 0.12) !important;
     }
-
     .gradio-textbox input::placeholder,
     .gradio-textbox textarea::placeholder {
         color: rgba(255, 255, 255, 0.5) !important;
     }
-
     /* Labels */
     .gradio-container label {
         color: var(--text-secondary) !important;
@@ -651,7 +639,6 @@ def main():
         margin-bottom: 6px !important;
         display: block !important;
     }
-
     /* Buttons */
     .gradio-container button {
         background: linear-gradient(135deg, var(--accent-purple) 0%, var(--accent-blue) 100%) !important;
@@ -669,339 +656,11 @@ def main():
         align-items: center !important;
         justify-content: center !important;
     }
-
     .gradio-container button:hover {
         transform: translateY(-2px) !important;
         box-shadow: 0 8px 25px rgba(139, 92, 246, 0.6) !important;
         background: linear-gradient(135deg, #9333ea 0%, #2563eb 100%) !important;
     }
-
     .gradio-container button:active {
         transform: translateY(0px) !important;
-    }
-
-    /* Send button specific styling */
-    .gradio-container .gr-button {
-        background: linear-gradient(135deg, var(--accent-purple) 0%, var(--accent-blue) 100%) !important;
-        border: 1px solid rgba(255, 255, 255, 0.2) !important;
-        color: white !important;
-        font-weight: 600 !important;
-        text-transform: none !important;
-        letter-spacing: 0.5px !important;
-    }
-
-    /* Chatbot specific styling */
-    .gradio-chatbot {
-        background: transparent !important;
-        border: none !important;
-    }
-
-    /* Footer */
-    .footer {
-        text-align: center !important;
-        padding: 28px !important;
-        color: var(--text-secondary) !important;
-        background: var(--glass-bg) !important;
-        backdrop-filter: blur(20px) !important;
-        border: 1px solid var(--glass-border) !important;
-        border-radius: 20px !important;
-        margin-top: 32px !important;
-        box-shadow: var(--shadow-glow) !important;
-    }
-
-    /* Scrollbar styling */
-    ::-webkit-scrollbar {
-        width: 8px;
-    }
-
-    ::-webkit-scrollbar-track {
-        background: rgba(255, 255, 255, 0.05);
-        border-radius: 4px;
-    }
-
-    ::-webkit-scrollbar-thumb {
-        background: linear-gradient(135deg, var(--accent-purple), var(--accent-blue));
-        border-radius: 4px;
-    }
-
-    ::-webkit-scrollbar-thumb:hover {
-        background: linear-gradient(135deg, #9333ea, #2563eb);
-    }
-
-    /* Responsive design */
-    @media (max-width: 768px) {
-        .gradio-container {
-            padding: 12px !important;
-        }
-        
-        .header {
-            padding: 20px !important;
-            border-radius: 16px !important;
-        }
-        
-        .header h1 {
-            font-size: 28px !important;
-        }
-        
-        .tools-area,
-        .chat-area {
-            border-radius: 16px !important;
-            padding: 20px !important;
-        }
-    }
-    """
-    
-    # Create wrapper function for proper chat handling
-    def respond(message, history):
-        """Response handler for Gradio messages format"""
-        if not message.strip():
-            return "", history
-        
-        # Convert messages format to simple tuples
-        simple_history = []
-        for i in range(0, len(history), 2):
-            if i + 1 < len(history):
-                user_msg = history[i].get('content', '') if isinstance(history[i], dict) else str(history[i])
-                bot_msg = history[i + 1].get('content', '') if isinstance(history[i + 1], dict) else str(history[i + 1])
-                if user_msg and bot_msg:
-                    simple_history.append([user_msg, bot_msg])
-        
-        # Generate response
-        for response_chunk in creed_ai.generate_response(message, simple_history):
-            # Create new history with the streaming response
-            new_history = history + [
-                {"role": "user", "content": message},
-                {"role": "assistant", "content": response_chunk}
-            ]
-            yield "", new_history
-    
-    # Create the interface with modern theme
-    with gr.Blocks(
-        title="🎸 Creed Bratton AI",
-        css=modern_css,
-        theme=gr.themes.Base()  # Use base theme for better CSS control
-    ) as demo:
-        
-        # Modern header
-        gr.HTML(f"""
-        <div class="header">
-            <h1>🎸 Creed Bratton AI</h1>
-            <p>Powered by phxdev/creed-qwen-0.5b-lora • Running on {'⚡ ZeroGPU' if SPACES_AVAILABLE else '🖥️ CPU'}</p>
-        </div>
-        """)
-        
-        # Model info with glass styling
-        gr.HTML("""
-        <div class="info-box">
-            <strong>Model:</strong> phxdev/creed-qwen-0.5b-lora<br>
-            <strong>Base:</strong> Qwen 0.5B + LoRA fine-tuning<br>
-            <strong>Tokens:</strong> &lt;thinking&gt;, &lt;conspiracy&gt;, &lt;tangent&gt;<br>
-            <strong>Mode:</strong> ZeroGPU optimized + Coherence validation
-        </div>
-        """)
-        
-        # MCP status
-        if os.environ.get('GRADIO_MCP_ENABLED'):
-            gr.HTML("""
-            <div class="status-box">
-                ✓ MCP Server Active • Available as tool for Claude Desktop
-            </div>
-            """)
-        
-        # Main chat interface with glass styling
-        with gr.Row(elem_classes="chat-area"):
-            chatbot = gr.Chatbot(
-                type='messages',  # Use messages format (modern)
-                height=550,
-                show_copy_button=True,
-                show_share_button=False,
-                avatar_images=["👤", "🎸"],
-                bubble_full_width=False,
-                show_label=False,
-                placeholder="🎸 Creed is ready...",
-                container=False
-            )
-        
-        # Input with explicit send button
-        with gr.Row():
-            with gr.Column(scale=7):
-                msg = gr.Textbox(
-                    placeholder="Ask Creed anything...",
-                    container=False,
-                    submit_btn=False,  # Disable built-in submit
-                    stop_btn=False
-                )
-            with gr.Column(scale=1, min_width=100):
-                send_btn = gr.Button("Send", variant="primary", size="lg")
-        
-        # Wire up the chat - both Enter key and Send button
-        msg.submit(
-            respond,
-            inputs=[msg, chatbot],
-            outputs=[msg, chatbot],
-            show_progress="hidden"
-        )
-        
-        send_btn.click(
-            respond,
-            inputs=[msg, chatbot],
-            outputs=[msg, chatbot],
-            show_progress="hidden"
-        )
-        
-        # MCP Tools section with glass styling
-        with gr.Row(elem_classes="tools-area"):
-            gr.HTML('<div class="tools-title">🛠️ Claude → Creed Transformation Tools</div>')
-            gr.HTML("""
-            <div style="margin-bottom: 20px; padding: 12px; background: rgba(255,255,255,0.1); border-radius: 8px; font-size: 14px;">
-                <strong>These MCP tools transform Claude into Creed Bratton for your Claude Desktop workflow.</strong><br>
-                Test them here, then use them in Claude Desktop to get Creed-mode responses from Claude.
-            </div>
-            """)
-            
-            with gr.Row():
-                with gr.Column():
-                    wisdom_topic = gr.Textbox(
-                        label="Wisdom Topic",
-                        placeholder="life, business, relationships..."
-                    )
-                    wisdom_output = gr.Textbox(
-                        label="Creed Wisdom Prompt (Copy to Claude Desktop)", 
-                        interactive=False,
-                        lines=5
-                    )
-                    wisdom_btn = gr.Button("Generate Wisdom Prompt", variant="primary")
-                    
-                with gr.Column():
-                    story_situation = gr.Textbox(
-                        label="Story Request",
-                        placeholder="mysterious, business, quarry..."
-                    )
-                    story_output = gr.Textbox(
-                        label="Creed Story Prompt (Copy to Claude Desktop)",
-                        interactive=False,
-                        lines=5
-                    )
-                    story_btn = gr.Button("Generate Story Prompt", variant="primary")
-            
-            with gr.Row():
-                with gr.Column():
-                    analysis_subject = gr.Textbox(
-                        label="Analysis Subject",
-                        placeholder="business strategy, market trends..."
-                    )
-                    analysis_output = gr.Textbox(
-                        label="Creed Analysis Prompt",
-                        interactive=False,
-                        lines=4
-                    )
-                    analysis_btn = gr.Button("Generate Analysis Prompt", variant="primary")
-                    
-                with gr.Column():
-                    conspiracy_topic = gr.Textbox(
-                        label="Conspiracy Topic", 
-                        placeholder="government, corporations, technology..."
-                    )
-                    conspiracy_output = gr.Textbox(
-                        label="Creed Conspiracy Prompt",
-                        interactive=False,
-                        lines=4
-                    )
-                    conspiracy_btn = gr.Button("Generate Conspiracy Prompt", variant="primary")
-                    
-            with gr.Row():
-                activate_output = gr.Textbox(
-                    label="Creed Mode Activation Prompt (Full Transformation)",
-                    interactive=False,
-                    lines=6
-                )
-                activate_btn = gr.Button("🎸 Generate FULL Creed Mode Prompt", variant="secondary", size="lg")
-            gr.HTML('<div class="tools-title">🛠️ MCP Tools</div>')
-            
-            with gr.Row():
-                with gr.Column():
-                    wisdom_topic = gr.Textbox(
-                        label="Wisdom Topic",
-                        placeholder="life, business, relationships..."
-                    )
-                    wisdom_output = gr.Textbox(
-                        label="Creed's Response", 
-                        interactive=False,
-                        lines=3
-                    )
-                    wisdom_btn = gr.Button("Ask Creed", variant="primary")
-                    
-                with gr.Column():
-                    story_situation = gr.Textbox(
-                        label="Story Request",
-                        placeholder="Tell me about..."
-                    )
-                    story_output = gr.Textbox(
-                        label="Creed's Story",
-                        interactive=False,
-                        lines=3
-                    )
-                    story_btn = gr.Button("Get Story", variant="primary")
-        
-        # Wire up the tools - now they're Claude transformation tools
-        wisdom_btn.click(
-            lambda topic: creed_ai.creed_wisdom_tool(topic or "life"),
-            inputs=[wisdom_topic],
-            outputs=[wisdom_output]
-        )
-        
-        story_btn.click(
-            lambda situation: creed_ai.creed_story_tool(situation or "mysterious"),
-            inputs=[story_situation],
-            outputs=[story_output]
-        )
-        
-        # Modern footer
-        gr.HTML("""
-        <div class="footer">
-            <strong>Creed Bratton AI</strong><br>
-            Model: phxdev/creed-qwen-0.5b-lora • Trained by Mark Scott<br>
-            <em>"Sometimes a guy's gotta ride the bull, am I right?"</em>
-        </div>
-        """)
-    
-    # Launch with modern styling and public sharing
-    print("🚀 Launching Real Creed AI with modern glassmorphism design...")
-    def _get_fallback_response(self) -> str:
-    """Fallback response when all generation methods fail"""
-    fallback_responses = [
-        "🎸 *Creed stares blankly* You know what? Let me think about that one...",
-        "🎸 That reminds me of something, but I can't quite remember what. Want to try asking again?",
-        "🎸 *Creed looks confused* Sometimes my brain just... goes places. What were we talking about?",
-        "🎸 You know, back in the '60s, we had a saying... actually, I forgot what it was. Try me again.",
-        "🎸 *Creed scratches his head* I'm getting one of those feelings... like when you forget why you walked into a room.",
-    ]
-    import random
-    return random.choice(fallback_responses)
-
-    def _is_coherent(self, response: str) -> bool:
-        """Check if response is coherent enough"""
-        if not response or len(response.strip()) < 5:
-            return False
-        
-        # Basic coherence checks
-        if response.count("Human:") > 2 or response.count("Assistant:") > 2:
-            return False
-        
-        # Check for obvious model artifacts
-        artifacts = ["<|im_end|>", "<|im_start|>", "[INST]", "[/INST]"]
-        if any(artifact in response for artifact in artifacts):
-            return False
-        
-        return True
-    demo.launch(
-        ssr_mode=False,
-        server_name="0.0.0.0",
-        server_port=7860,
-        share=True,  # Create public link
-        show_error=True,
-        mcp_server=True
-    )
-
-if __name__ == "__main__":
-    main()
+    }