Remove mistral_common to fix dependency conflict

- Remove mistral_common from requirements (conflicts with numpy/pydantic)
- Simplify prompt_formatter to use manual Mistral format
- Format: [INST] {system}\n\n{user} [/INST] (no <s>, tokenizer adds BOS)
- Keep recommended_temperature=0.15 for Devstral

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <[email protected]>

backend/prompt_formatter.py

@@ -5,54 +5,12 @@ Handles formatting prompts appropriately for different model types:
 - Instruction models: System prompt + user message with chat template
 """
 
-from typing import Dict, Optional, Any, List
+from typing import Dict, Optional, Any
 import logging
 
 logger = logging.getLogger(__name__)
 
 
-def _try_mistral_common_format(messages: List[Dict[str, str]], model_name: str) -> Optional[str]:
-    """
-    Try to use mistral_common for proper Mistral/Devstral chat formatting.
-    Returns None if mistral_common is not available or fails.
-    """
-    try:
-        from mistral_common.protocol.instruct.messages import (
-            SystemMessage, UserMessage
-        )
-        from mistral_common.protocol.instruct.request import ChatCompletionRequest
-        from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
-
-        # Load the tokenizer from HF hub
-        tokenizer = MistralTokenizer.from_hf_hub(model_name)
-
-        # Build messages
-        mistral_messages = []
-        for msg in messages:
-            if msg["role"] == "system":
-                mistral_messages.append(SystemMessage(content=msg["content"]))
-            elif msg["role"] == "user":
-                mistral_messages.append(UserMessage(content=msg["content"]))
-
-        # Encode to get token IDs
-        request = ChatCompletionRequest(messages=mistral_messages)
-        tokenized = tokenizer.encode_chat_completion(request)
-
-        # Decode back to text for use with HF tokenizer
-        # This gives us the properly formatted prompt string
-        decoded = tokenizer.decode(tokenized.tokens)
-
-        logger.info(f"Used mistral_common format for {model_name}")
-        return decoded
-
-    except ImportError:
-        logger.warning("mistral_common not available, using fallback format")
-        return None
-    except Exception as e:
-        logger.warning(f"mistral_common formatting failed: {e}, using fallback")
-        return None
-
-
 class PromptFormatter:
     """
     Unified prompt formatting for different model types.
@@ -62,9 +20,9 @@ class PromptFormatter:
         - Model treats it as text to continue
 
     Instruction models (Devstral, instruct variants):
-        - Use mistral_common for Mistral/Devstral models
-        - Fallback to tokenizer's chat_template if available
-        - Final fallback to manual Mistral format
+        - Wrap with system prompt + user message
+        - Use tokenizer's chat_template if available
+        - Fallback to manual Mistral format
     """
 
     def format(
@@ -110,9 +68,8 @@ class PromptFormatter:
         Format prompt for instruction-tuned models.
 
         Priority:
-        1. mistral_common for Mistral/Devstral models
-        2. Tokenizer's native chat_template
-        3. Manual Mistral format fallback
+        1. Tokenizer's native chat_template (if available)
+        2. Manual Mistral format fallback
         """
         # Get system prompt (override > model default > generic fallback)
         system_prompt = system_prompt_override or model_config.get("system_prompt")
@@ -125,15 +82,7 @@ class PromptFormatter:
             {"role": "user", "content": prompt}
         ]
 
-        # For Mistral/Devstral models, try mistral_common first
-        architecture = model_config.get("architecture", "")
-        hf_path = model_config.get("hf_path", "")
-        if architecture == "mistral" or "mistral" in hf_path.lower():
-            formatted = _try_mistral_common_format(messages, hf_path)
-            if formatted:
-                return formatted
-
-        # Try tokenizer's native chat template
+        # Try tokenizer's native chat template first
         if hasattr(tokenizer, 'chat_template') and tokenizer.chat_template is not None:
             try:
                 formatted = tokenizer.apply_chat_template(
@@ -146,16 +95,18 @@ class PromptFormatter:
             except Exception as e:
                 logger.warning(f"chat_template failed: {e}, using manual format")
 
-        # Fallback: Manual Mistral/Llama format
-        # Note: Don't include <s> as the tokenizer adds it during tokenization
+        # Fallback: Manual Mistral/Llama instruction format
+        # Note: We DON'T include <s> - the tokenizer adds BOS automatically
         return self._manual_mistral_format(prompt, system_prompt)
 
     def _manual_mistral_format(self, prompt: str, system_prompt: str) -> str:
         """
-        Manual Mistral instruction format as fallback.
+        Manual Mistral instruction format.
 
         Format: [INST] {system}\n\n{user} [/INST]
-        Note: <s> is NOT included as the tokenizer adds BOS automatically.
+
+        Note: BOS token (<s>) is NOT included - the tokenizer adds it
+        automatically during tokenization with add_special_tokens=True (default).
         """
         logger.info("Using manual Mistral instruction format")
         return f"[INST] {system_prompt}\n\n{prompt} [/INST]"

requirements.txt

@@ -10,7 +10,6 @@ pydantic==2.5.0
 torch>=2.3.0
 transformers>=4.44.0
 accelerate>=0.30.0
-mistral_common>=1.5.0  # Required for Devstral chat template formatting
 
 # Utilities
 numpy==1.24.3