Statistics Netherlands Logo UvA Logo INDElab Logo

🦗 LOCuST: Large-scale Official statistics Complex text-to-SQL Task

1Statistics Netherlands, 2University of Amsterdam

About LOCuST

LOCuST is a large-scale text-to-SQL benchmark on real-world statistical data provided by Statistics Netherlands. Our benchmark contains over 2,000 English and Dutch tables covering 22 statistical domains. In concrete, our datasets provide the following:

  • 2,244 tables containing real-world statistics;
  • table schemas ranging from small (< 5 columns) to very large (> 1,200 columns);
  • a coverage of 22 statistical domains;
  • 2,567 manually annotated complex questions and answers;
  • an extensive metadata knowledge graph accompanying the tables, containing labels, descriptions, units, and other associated metadata for covered concepts.
Our benchmark consists of three tasks: (1) table retrieval; (2) query generation, where tables are given; and (3) end-to-end QA, which is a combination of table retrieval and query generation. All tasks are run on the same question-answer data.

Setup

The total size of the datasets when extracted are 2.03 GB (en) and 4.68 GB (nl)

The knowledge graphs files can be downloaded from here: For running models and evaluating your own solutions, please refer to the setup guide in the code repository here.

Changelog

Citation

If you use LOCuST in your research, please cite our work:

@article{lageweg2026locust,
    title={A Text-to-SQL Benchmark for Official Statistics},
    author={}
    year={2026}
}

Leaderboard (English)

🇬🇧

Leaderboard (Dutch)

🇳🇱