changed LLM again

app.py

@@ -357,7 +357,7 @@ class CVEDashboard:
 
 def generate_tailored_summary(cve_description: str, audience: str, hf_token: Optional[str] = None, max_retries: int = 2) -> str:
     """
-    Generates a tailored CVE summary using Apertus via HuggingFace Inference API.
+    Generates a tailored CVE summary using SmolLM3-3B via HuggingFace Inference API.
     
     Args:
         cve_description: The original CVE description
@@ -380,20 +380,19 @@ def generate_tailored_summary(cve_description: str, audience: str, hf_token: Opt
     if audience not in AUDIENCE_PROFILES:
         return f"❌ Unknown audience: {audience}"
     
-    # Try Apertus models with fallback options
+    # Define the model to use
     models = [
-        "swiss-ai/Apertus-8B-Instruct-2509",  # Primary choice - smaller, faster
-        "mistralai/Mistral-7B-Instruct-v0.3"  # Fallback to original
+        "HuggingFaceTB/SmolLM3-3B",
     ]
     
     headers = {"Authorization": f"Bearer {token}"}
     profile = AUDIENCE_PROFILES[audience]
     
-    # Apertus uses chat template format
-    messages = [
-        {
-            "role": "user", 
-            "content": f"""You are an expert cybersecurity analyst. Rewrite this CVE description for a {audience}.
+    # SmolLM3 uses a specific chat template format.
+    # Build the prompt string manually based on its tokenizer_config.json.
+    # The system prompt includes /no_think to get a direct answer without a reasoning trace.
+    system_prompt = "You are an expert cybersecurity analyst. /no_think"
+    user_prompt = f"""Rewrite this CVE description for a {audience}.
 **Target Audience:** {audience}
 **Focus:** {profile['focus']}
 **Tone:** {profile['tone']}
@@ -401,65 +400,38 @@ def generate_tailored_summary(cve_description: str, audience: str, hf_token: Opt
 **CVE Description:**
 {cve_description[:1200]}
 Provide a concise, actionable summary (2-3 sentences) highlighting what matters most to this audience. Focus on practical implications and next steps."""
-        }
-    ]
+
+    # Manually apply the chat template
+    prompt = f"### System:\n{system_prompt}\n### User:\n{user_prompt}\n### Assistant:\n"
     
     for model in models:
         api_url = f"https://api-inference.huggingface.co/models/{model}"
         
-        # Different payload structure for Apertus vs Mistral
-        if "Apertus" in model:
-            payload = {
-                "inputs": {
-                    "messages": messages,
-                    "max_tokens": 200,
-                    "temperature": 0.8,  # Recommended by Apertus docs
-                    "top_p": 0.9        # Recommended by Apertus docs
-                }
-            }
-        else:
-            # Fallback to Mistral format
-            prompt = f"""<s>[INST] You are an expert cybersecurity analyst. Rewrite the following CVE description for a {audience}.
-**Focus:** {profile['focus']}
-**Tone:** {profile['tone']}
-CVE: {cve_description[:1000]}
-Provide a 2-3 sentence summary highlighting what matters most to this audience: [/INST]"""
-            
-            payload = {
-                "inputs": prompt,
-                "parameters": {
-                    "max_new_tokens": 150,
-                    "temperature": 0.5,
-                    "top_p": 0.9,
-                    "do_sample": True,
-                    "return_full_text": False,
-                    "stop": ["\n\n"]
-                }
+        payload = {
+            "inputs": prompt,
+            "parameters": {
+                "max_new_tokens": 200,
+                "temperature": 0.6,      # Recommended
+                "top_p": 0.95,           # Recommended
+                "do_sample": True,
+                "return_full_text": False,
+                "stop": ["\n###", "<|endoftext|>"] # Stop sequences
             }
+        }
         
         for attempt in range(max_retries):
             try:
                 logger.info(f"Generating summary with {model} (attempt {attempt + 1})")
                 
                 response = requests.post(api_url, headers=headers, json=payload, timeout=45)
-                
                 if response.status_code == 200:
                     try:
                         result = response.json()
                         
-                        # Handle different response formats
+                        # Standard response format for this payload type
                         summary = ""
-                        if "Apertus" in model:
-                            # Apertus response format
-                            if isinstance(result, list) and len(result) > 0:
-                                if "generated_text" in result[0]:
-                                    summary = result[0]["generated_text"]
-                                elif "choices" in result[0] and len(result[0]["choices"]) > 0:
-                                    summary = result[0]["choices"][0].get("message", {}).get("content", "")
-                        else:
-                            # Mistral response format
-                            if isinstance(result, list) and len(result) > 0:
-                                summary = result[0].get('generated_text', '').strip()
+                        if isinstance(result, list) and len(result) > 0:
+                            summary = result[0].get('generated_text', '').strip()
                         
                         if summary and len(summary) > 20:
                             logger.info(f"Successfully generated summary with {model}")
@@ -616,8 +588,7 @@ def create_interface():
                             - Search CVEs by date range and keywords
                             - Filter by severity levels
                             - Visualize CVE distributions and trends
-                            - Export data to JSON or CSV formats
-                            - **NEW:** AI-powered audience-specific summaries using multiple LLMs
+                            - AI-powered audience-specific summaries using the SmolLM3-3B model.
                             
                             **Supported Audiences:**
                             - **Cybersecurity Professional:** Focus on threats, attack vectors, and mitigation
@@ -629,11 +600,12 @@ def create_interface():
                             
                             **Data Source:** [NIST NVD API](https://nvd.nist.gov/developers/vulnerabilities)
                             
-                            **AI Models:** Multiple HuggingFace models with fallback support
+                            **AI Model:** [HuggingFaceTB/SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B)
+
+                            **Disclaimer:** Generated content may be inaccurate or false.
                             
                             **Performance Optimizations:**
                             - Shorter timeouts for faster failure detection
-                            - Multiple model fallback for reliability
                             - Optimized prompts for quicker responses
                             
                             **Rate Limits:**