TheJoin: A Collection of 650 Multi-Domain Relational Databases for Pretraining

A collection of 650 relational databases spanning domains like e-commerce, finance, sports, biomedical, and government, ported to the RelBench manifest format. It was created by stanford-rdl for large-scale pretraining of relational and tabular foundation models, with each database being self-describing and tasks shipping labels as-is.

Use Cases

• Pretraining relational foundation models based on the collection's 650 multi-domain databases.
• Benchmarking tabular machine learning models across diverse domains like e-commerce and finance mentioned in the description.
• Developing self-describing data systems based on the RelBench manifest format used by the collection.

Strengths

• Contains 650 distinct relational databases, indicating substantial scale.
• Spans many domains including academic, e-commerce, finance, sports, biomedical, and government, suggesting broad coverage.
• Databases are self-describing and formatted for pretraining, which likely aids in automated processing.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Freshness should be verified; the last metadata update was 2026-06-12.

Provenance

Source

stanford-rdl

Collection Method

Ported from various sources to the RelBench manifest format.

Freshness

Last updated 2026-06-12 19:54:50; freshness should be verified.