DB-GPT-Hub is an experimental Python project for Text-to-SQL using LLMs, with SFT fine-tuning and multiple model backbones. It provides a quickstart workflow and example configs for training, predicting, and evaluating.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
DB-GPT-Hub is an experimental project that leverages Large Language Models (LLMs) to achieve Text-to-SQL parsing. The project encompasses data collection, data preprocessing, model selection and construction, and fine-tuning of model weights to enhance Text-to-SQL capabilities while reducing training costs.
As of October 10, 2023, it reports fine-tuning an open-source 13B-sized model with Spider-based test-suite achieving 0.764 execution accuracy on a 1.27G database, and 0.825 on the Spider database (95M). The repository integrates multiple large models and includes a workflow for data processing, supervised fine-tuning, prediction output, and evaluation.
How it works
DB-GPT-Hub applies Supervised Fine-Tuning (SFT) on large language models. It supports a set of base models and uses Quantized Learning over Redundant Architecture (QLoRA) with a quantization bit of 4 for fine-tuning. It outlines hardware guidance and minimum memory requirements per model size.
Getting started
Environment and quick start steps are provided:
- Environment setup snippet:
git clone https://github.com/eosphoros-ai/DB-GPT-Hub.git
cd DB-GPT-Hub
conda create -n dbgpt_hub python=3.10
conda activate dbgpt_hub
cd src/dbgpt_hub_sql
pip install -e .
- Quick Start snippet:
pip install dbgpt-hub
Then a Python-based workflow is shown to preprocess data, start SFT training, run prediction, and evaluate.
Data preparation
The README describes using Spider as the primary dataset and mentions additional datasets such as WikiSQL, CHASE, BIRD-SQL, and CoSQL. It references a processing template yielding approximately 20W dataset.
Recent releases
Latest releases include:
- v0.3.0 (2023-12-08): Added more models and experiments/evaluation results, with expanded datasets and APIs.
- v0.2.0 (2023-09-25): Refactored project code and optimizations; improved Spider evaluation execution accuracy.
- v0.0.2 (2023-07-30): Added LoRA fine-tuning for Llama/Llama2, QLoRA for Llama2, multi-GPU support.
- v0.0.1 (2023-07-13): Initial scaffold with Spider + QLoRA + Falcon SFT and readme corrections.
Traction
Stars: 2004
Behind the repo
The repository is part of the DB-GPT-Hub project by eosphoros-ai and mentions integration with Spider/test-suite and various model families.
Caveats
License: MIT. Created 2023-06-02. Last push 2025-07-02. Open issues: 73. Language: Python.