Refactor settings.py: Update model configurations to replace OpenAssistant SFT-1 and SOLAR 10.7B with Yi 6B Chat and Mixtral 8x7B, enhancing descriptions and capabilities

config/settings.py

@@ -214,10 +214,10 @@ MODELS = {
             "documentation": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2"
         }
     },
-    "oasst-sft-1": {
-        "id": "OpenAssistant/oasst-sft-1-pythia-12b",
-        "name": "OpenAssistant SFT-1",
-        "description": "Open Assistant model trained for dialogue and instruction following",
+    "yi-6b": {
+        "id": "01-ai/Yi-6B-Chat",
+        "name": "Yi 6B Chat",
+        "description": "Lightweight multilingual model with strong performance",
         "type": "base",
         "parameters": {
             "max_length": 2048,
@@ -226,8 +226,8 @@ MODELS = {
             "repetition_penalty": 1.1,
         },
         "training": {
-            "base_model_path": "OpenAssistant/oasst-sft-1-pythia-12b",
-            "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "oasst-sft-1-tuned"),
+            "base_model_path": "01-ai/Yi-6B-Chat",
+            "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "yi-6b-chat-tuned"),
             "lora_config": {
                 "r": 16,
                 "lora_alpha": 32,
@@ -236,31 +236,31 @@ MODELS = {
             }
         },
         "details": {
-            "full_name": "Open Assistant SFT-1",
+            "full_name": "Yi 6B Chat",
             "capabilities": [
-                "Strong dialogue capabilities",
-                "Good multilingual performance",
-                "Instruction following",
-                "Community-trained on diverse datasets"
+                "Strong multilingual support",
+                "Efficient performance",
+                "Good instruction following",
+                "Smaller size (6B parameters)"
             ],
             "limitations": [
-                "May require more specific prompting for legal domain",
-                "Performance varies by language",
-                "Less specialized in legal terminology"
+                "Less specialized in legal domain",
+                "May need specific prompting",
+                "Newer model with less community testing"
             ],
             "use_cases": [
-                "Multilingual client support",
-                "General legal consultation",
+                "General consultation",
                 "Document analysis",
-                "Cross-lingual communication"
+                "Multilingual support",
+                "Basic legal QA"
             ],
-            "documentation": "https://huggingface.co/OpenAssistant/oasst-sft-1-pythia-12b"
+            "documentation": "https://huggingface.co/01-ai/Yi-6B-Chat"
         }
     },
-    "solar-10.7b": {
-        "id": "upstage/SOLAR-10.7B-Instruct-v1.0",
-        "name": "SOLAR 10.7B Instruct",
-        "description": "Lightweight multilingual model with strong Asian language support",
+    "mixtral-8x7b": {
+        "id": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+        "name": "Mixtral 8x7B Instruct",
+        "description": "Powerful mixture-of-experts model with strong multilingual capabilities",
         "type": "base",
         "parameters": {
             "max_length": 2048,
@@ -269,8 +269,8 @@ MODELS = {
             "repetition_penalty": 1.1,
         },
         "training": {
-            "base_model_path": "upstage/SOLAR-10.7B-Instruct-v1.0",
-            "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "solar-10.7b-tuned"),
+            "base_model_path": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "mixtral-8x7b-instruct-tuned"),
             "lora_config": {
                 "r": 16,
                 "lora_alpha": 32,
@@ -279,68 +279,25 @@ MODELS = {
             }
         },
         "details": {
-            "full_name": "SOLAR 10.7B Instruct v1.0",
+            "full_name": "Mixtral 8x7B Instruct v0.1",
             "capabilities": [
-                "Strong Asian language support",
-                "Excellent instruction following",
-                "Fast inference speed",
-                "Good multilingual performance"
+                "Excellent multilingual support",
+                "Strong reasoning capabilities",
+                "Superior instruction following",
+                "High-quality text generation"
             ],
             "limitations": [
-                "Slightly larger size (10.7B)",
-                "May need more specific prompting for legal domain",
-                "Performance varies by language"
+                "Larger model size",
+                "May need domain-specific prompting",
+                "Higher resource requirements"
             ],
             "use_cases": [
-                "Multilingual legal assistance",
-                "Asian language support",
-                "Cross-cultural legal consultation",
-                "International document analysis"
-            ],
-            "documentation": "https://huggingface.co/upstage/SOLAR-10.7B-Instruct-v1.0"
-        }
-    },
-    "saiga-7b": {
-        "id": "IlyaGusev/saiga_7b_lora",
-        "name": "Saiga 7B",
-        "description": "Multilingual model optimized for Russian and English dialogue",
-        "type": "base",
-        "parameters": {
-            "max_length": 2048,
-            "temperature": 0.7,
-            "top_p": 0.9,
-            "repetition_penalty": 1.1,
-        },
-        "training": {
-            "base_model_path": "IlyaGusev/saiga_7b_lora",
-            "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "saiga-7b-tuned"),
-            "lora_config": {
-                "r": 16,
-                "lora_alpha": 32,
-                "lora_dropout": 0.05,
-                "target_modules": ["q_proj", "v_proj", "k_proj", "o_proj"]
-            }
-        },
-        "details": {
-            "full_name": "Saiga 7B",
-            "capabilities": [
-                "Excellent Russian language support",
-                "Strong dialogue capabilities",
-                "Optimized for instruction following",
-                "Good performance on legal and technical texts"
-            ],
-            "limitations": [
-                "Primarily focused on Russian and English",
-                "May require specific prompting for legal domain",
-                "Limited support for other languages"
-            ],
-            "use_cases": [
-                "Russian-English legal consultation",
-                "Document analysis in Russian",
-                "Legal QA for Russian speakers",
-                "Technical documentation processing"
+                "Complex legal analysis",
+                "Multilingual consultation",
+                "Advanced document processing",
+                "Cross-lingual communication"
             ],
-            "documentation": "https://huggingface.co/IlyaGusev/saiga_7b_lora"
+            "documentation": "https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1"
         }
     }
 }