GSGrounded SQL
Hugging Face GitHub ↗

text-to-sql · qlora fine-tune · execution-verified

Ask a database a question in plain English.

Grounded SQL fine-tunes Mistral-7B-Instruct with QLoRA, narrows the schema with hybrid retrieval before it generates anything, and checks every answer by actually running the SQL against the database, not by guessing whether it looks right.

See it work below ↓ See the full architecture →
70.2%execution accuracy
0%hallucinated columns
340msp95 latency, INT8
r=64QLoRA rank
live example

Ask in plain English

This hosted preview replays real, cached examples against a live SQLite database. It isn't running the full 7B model in your browser. To generate SQL for any question against any schema, run Grounded SQL locally →.

Tables pulled by hybrid retrieval

Generated SQL

schemas in play

The demo above queries these real databases

Two schemas from the Spider benchmark, switchable above, covering different domains and table shapes, executed against their real .sqlite files.

Seven tables total across concert_singer and pets_1. Switch the tab in the demo above to try either one.

go deeper

Three things worth understanding

This page shows the product. These go into how it's actually built.

01 → 04

How it works

Four stages sit between your question and a trustworthy answer: retrieval narrows the schema, a fine-tuned model writes the SQL, and execution proves it's correct.

Read the architecture → 70.2%

Results

A four-way comparison against the base model and two RAG variants, scored on 1,034 real test queries by actually executing them.

See the full benchmark → bring your own data

Run it yourself

Clone the repo, download the dataset, get the adapter, and point the Gradio UI at your own schema or .sqlite file.

Get started →