🔥 Natural Language to SQL Generation using Open-Source LLMs

A two-stage generative AI system that converts natural language questions into SQL queries, using schema-linking and fine-tuned Large Language Models (LLMs). Built during my final-year project at AxeFinance.

🚀 Features

• Two-stage architecture:
  • Schema Linking Model: Identifies correct tables & columns.
  • SQL Generation Model: Generates SQL queries.
• Fine-tuned open-source models (Mistral, LLaMA).
• API endpoint built using FastAPI.
• Dataset: Spider benchmark.
• Achieved 83.46% execution accuracy on Spider.

📷 Demo

📦 Technologies Used

• Python
• Hugging Face Transformers
• PEFT / LoRA
• PyTorch
• FastAPI
• SQL Server (for testing)
• Google Colab (for training)

📊 Results

• Execution Accuracy: 83.46%
• Schema Linking Accuracy: 91.66%
• Ranked 3rd on Spider Benchmark

🔗 Models on Hugging Face

• Schema-Linking Model
• SQL Generation Model

🚀 How to Run

pip install -r requirements.txt
uvicorn src.api:app --reload