How boostcampaitech5/level3_nlp_finalproject-nlp-12 works
Main Components
a. Data Pipeline Components:
• Data Generation & Processing Layer
• Synthetic data generation flow
• Data preprocessing pipelines
• EDA components
• Both single-turn and multi-turn dialogue processing
b. Model Architecture:
• Base Model Layer (KULLM-Polyglot-Ko-12.8B)
• QLoRA Fine-tuning Layer
• Model Training Pipeline
• Inference Pipeline
c. Server Architecture:
• Frontend (React-based)
• Backend (Python/Flask)
• API Layer
• Database Connection
Component Relationships
[Data Processing] → [Model Layer] → [Server Layer] → [Client Layer]
Key flows to highlight:
Data Flow
• Raw data → Preprocessing → Training data
• Training data → Fine-tuning → Trained model
• User input → Model inference → Response
Service Flow
• Client request → Backend API → Model inference → Response formatting → Client display
Specific Implementation Details
Technical elements:
• QLoRA implementation with 4-bit quantization
• LangChain integration points
• Prompt template processing
• Evaluation pipeline (G-Eval and Human Evaluation)
Detailed Component Breakdown
Data Layer:
• Data Generation Scripts
• Preprocessing Pipeline
• EDA Tools
• Prompt Templates
Model Layer:
• KULLM-Polyglot-Ko-12.8B Base
• QLoRA Fine-tuning
• LangChain Integration
• Inference Engine
Server Layer:
• Flask Backend
• API Endpoints
• Database Connector
• Response Formatter
Client Layer:
• React Frontend
• Chat Interface
• State Management
• User Input Handler
Additional Elements
Supporting components:
• Configuration management
• Environment settings
• Evaluation pipelines
• Monitoring and logging
Key Technical Specifications
• Model parameters (12.8B)
• QLoRA specifications
• API endpoints
• Data volume indicators
Remember to use the project's actual directory structure to inform the component relationships, particularly the separation between finetuning, server, and data processing components as shown in the file tree.
This system design should help developers understand both the high-level architecture and the specific implementation details of the MyMary chatbot service.