DB-GPT-Hub

Why we included this project

Anyone tuning an open model to turn natural-language questions into SQL spends most of the time on setup: gathering datasets, writing training scripts, and wiring up evaluation. DB-GPT-Hub packages that whole supervised fine-tuning workflow, centered on the Spider benchmark and also covering WikiSQL, CHASE, and BIRD-SQL. It ships prepared data, LoRA/QLoRA training scripts for several 7B-13B open models, and an execution-accuracy evaluation stage that reports scores per difficulty tier. What you get is a repeatable pipeline from data preprocessing through training, prediction, and eval, so you can reproduce the published results and then point the same steps at your own schemas instead of trusting a single pre-trained checkpoint. Text2NLU and Text2GQL fine-tuning are covered as well, which helps if you are building conversational database agents. It is a solid reference workflow to copy before investing in your own Text-to-SQL experiments.