SQL3M: Token Efficient Text-to-SQL Generation

Figshare 2026 Dataset Open access

Read full text
Authors: Ibrahim Ücelehan, Alina Geiger, Dominik Sobania
DOI: 10.6084/m9.figshare.30178501
Cited by: 0

Abstract

This repository implements a two-stage retrieval system for Text-to-SQL tasks called SQL3M. It combines SBERT-based semantic retrieval with graph expansion via all shortest paths to obtain a structurally complete, compact database schema, reducing the number of tokens sent to large language models while preserving execution accuracy.

Fields of Study & Subjects

Topic Modeling Advanced Database Systems and Queries Natural Language Processing Techniques

Publication details

Journal / source
Figshare
Publication date
1 January 2026
Document type
Dataset