DTS-SQL

Overview

This repository hosts the code for the DTS-SQL paper, featuring state-of-the-art text-to-SQL models with 7B parameters. Our models demonstrate exceptional performance on the Spider and Spider-SYN datasets, setting new benchmarks in the field.

Repository Contents

Finetuning Scripts

• schema_linking_generation_finetuning.ipynb: Contains code for finetuning DeepSeek and Mistral models for text-to-SQL tasks.
• sql_generation_finetuning.ipynb: Dedicated to finetuning processes, specifically focusing on SQL generation.

Inference Scripts

• schema_linking_generation_inference.ipynb: Script for schema linking generation using the finetuned models.
• sql_generation_inference.ipynb: Script for SQL generation at inference time using the finetuned models.

Dataset Preparation

• finetuning_dataset_creator.py: Code for creating the finetuning dataset from the Spider dataset.

Utilities

• Utils directory: Contains all necessary helper functions for formatting tables and creating the dataset.

Results

• All of the results are stored in the results directory.

Citation

@article{pourreza2024dts, title={DTS-SQL: Decomposed Text-to-SQL with Small Large Language Models}, author={Pourreza, Mohammadreza and Rafiei, Davood}, journal={arXiv preprint arXiv:2402.01117}, year={2024} }