Lang2Query: Text-to-SQL Converter

📌 Overview

Lang2Query is an advanced Text-to-SQL translation system that enables users to query relational databases using natural language. It leverages state-of-the-art language modeling techniques to generate syntactically correct SQL queries, bridging the gap between intuitive user inputs and structured database retrieval.

🚀 Features

• Fine-Tuned Mistral Model: Optimized for SQL query generation using Low-Rank Adaptation (LoRA) and Retrieval-Augmented Generation (RAG).
• Multiple Baseline Models: Includes N-Gram, Seq2Seq LSTM, and Pre-trained Mistral for comparative evaluation.
• Efficient Data Preprocessing: Utilizes FAISS-based vector retrieval and UnSloth library for optimized tokenization.
• Evaluation Metrics: Assessed using ROUGE-1, ROUGE-L, BLEU, and BERTScore.

📊 Model Performance

Model	ROUGE-1	ROUGE-L	BLEU	BERTScore
N-Gram	0.060	0.050	0.0001	0.7800
Seq2Seq LSTM	0.090	0.090	0.0030	0.7900
Baseline Mistral	0.130	0.130	0.0070	0.8000
Fine-Tuned Mistral	0.230	0.200	0.0230	0.8100

📚 Dataset

• Source: Clinton/Text-to-SQL-v1 (HuggingFace)
• Size: 262,208 entries (~635 MB)
• Train-Validation-Test Split: 80%-10%-10%

🛠️ Data Processing & Model Training

• Vector Store Setup: FAISS-based indexing for similarity search.
• LoRA & RAG Integration: Enhances model adaptability to unseen SQL structures.
• Quantization: 4-bit weight compression for memory efficiency.
• Training Configuration:
  • Batch Size: 4 (Gradient Accumulation: 8)
  • Learning Rate: 0.0001
  • Precision: FP16 (Half-Precision)
  • Checkpoints: Saved every 50 steps (max 2 checkpoints)

🏆 Sample Prediction

Input Query	"What is the listed date of the Warren pony truss bridge built in 1925?"
Database Schema	table_8098 (Name, Built, Listed, Location, County, Type)
Ground Truth SQL	SELECT "Listed" FROM table_8098 WHERE "Built" = '1925' AND "Type" = 'warren pony truss'
Model Prediction	SELECT "Listed" FROM table_8098 WHERE "Incumbent" = 'john randolph redistricted from the 1'
BLEU Score	0.0245
ROUGE-L Score	0.3077

📜 Installation & Usage

🔧 Prerequisites

• Python 3.8+
• Install dependencies:

  pip install -r requirements.txt
  

🚀 Running the Notebook

1️⃣ Clone the repository git clone https://github.com/oberoiharshith/Lang2Query.git cd Lang2Query

2️⃣ Start Jupyter Notebook Open Lang2Query.ipynb and run the cells step by step.

📜 References

• TypeSQL (Yu et al., 2018)
• Dataset Decomposition for LLMs (Pouransari et al., 2024)
• Mistral-7B-Instruct