Refactor settings.py: Replace Phi-2 model configuration with Neural Mistral 7B, enhancing reasoning and instruction following capabilities

README.md

@@ -40,10 +40,10 @@ Status Law Assistant is a smart chatbot that answers user questions about Status
 - Model switching system with automatic fallback
 - Fine-tuning capabilities based on chat history
 - Multiple model support:
-  - Llama 2 7B Chat (primary): Optimized for dialogues
   - Zephyr 7B: Enhanced performance and response quality
-  - Mistral 7B Instruct v0.2: Superior multilingual capabilities
-  - XGLM 7.5B: Specialized cross-lingual generation model (requires paid API access)
+  - TinyLlama 1.1B Chat: Lightweight model for resource-constrained environments
+  - Neural Mistral 7B: Superior reasoning and instruction following capabilities
+  - Mixtral 8x7B: Advanced mixture-of-experts architecture
 
 ## 🚀 Technologies
 
@@ -135,17 +135,17 @@ The fine-tuning process uses LoRA (Low-Rank Adaptation) for efficient training w
 
 The application supports multiple models with automatic fallback:
 
-- Llama 2 7B Chat (default): Optimized for dialogues
 - Zephyr 7B: Enhanced performance and response quality
-- Mistral 7B Instruct v0.2: Superior multilingual capabilities
-- XGLM 7.5B: Specialized cross-lingual generation model (requires paid API access)
+- TinyLlama 1.1B Chat: Lightweight model for resource-constrained environments
+- Neural Mistral 7B: Superior reasoning and instruction following capabilities
+- Mixtral 8x7B: Advanced mixture-of-experts architecture
 
 Models can be switched dynamically through the interface or programmatically:
 
 ```python
 from src.training.model_manager import switch_to_model
 
-switch_to_model("llama-7b")  # or "zephyr-7b", "mistral-7b", "xglm-7b"
+switch_to_model("zephyr-7b")  # or "tinyllama-1.1b", "neural-mistral-7b", "mixtral-8x7b"
 ```
 
 ## 🔄 Knowledge Base Management

config/settings.py

@@ -217,10 +217,10 @@ MODELS = {
             "documentation": "https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1"
         }
     },
-    "phi-2": {  
-        "id": "microsoft/phi-2",
-        "name": "Phi-2",
-        "description": "Compact yet powerful 2.7B model with strong reasoning capabilities",
+    "neural-mistral": {  # заменяем phi-2
+        "id": "teknium/Neural-Mistral-7B-v0.1",
+        "name": "Neural Mistral 7B",
+        "description": "Enhanced version of Mistral with improved reasoning and instruction following",
         "type": "base",
         "parameters": {
             "max_length": 2048,
@@ -229,8 +229,8 @@ MODELS = {
             "repetition_penalty": 1.1,
         },
         "training": {
-            "base_model_path": "microsoft/phi-2",
-            "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "phi-2-tuned"),
+            "base_model_path": "teknium/Neural-Mistral-7B-v0.1",
+            "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "neural-mistral-7b-tuned"),
             "lora_config": {
                 "r": 16,
                 "lora_alpha": 32,
@@ -239,27 +239,28 @@ MODELS = {
             }
         },
         "details": {
-            "full_name": "Microsoft Phi-2",
+            "full_name": "Neural Mistral 7B v0.1",
             "capabilities": [
-                "Strong reasoning abilities",
-                "Excellent code understanding",
-                "Compact size (2.7B parameters)",
-                "Good performance-to-size ratio",
-                "Efficient resource usage",
-                "Research and commercial use allowed"
+                "Enhanced reasoning capabilities",
+                "Improved instruction following",
+                "Strong multilingual support",
+                "Better context understanding",
+                "Advanced problem-solving abilities",
+                "Consistent output quality"
             ],
             "limitations": [
-                "Smaller context window than larger models",
-                "Less specialized in legal domain",
-                "Limited multilingual capabilities"
+                "Requires more GPU memory",
+                "May be slower than smaller models",
+                "Resource intensive for fine-tuning"
             ],
             "use_cases": [
-                "Quick legal consultations",
-                "Document analysis",
-                "Code-related legal questions",
-                "Resource-efficient deployments"
+                "Complex legal analysis",
+                "Advanced reasoning tasks",
+                "Detailed document processing",
+                "Professional consultation",
+                "Research assistance"
             ],
-            "documentation": "https://huggingface.co/microsoft/phi-2"
+            "documentation": "https://huggingface.co/teknium/Neural-Mistral-7B-v0.1"
         }
     }
 }