RelBench v2: A Large-Scale Benchmark and Repository for Relational Data
Paper • 2602.12606 • Published
name stringclasses 3
values | domain stringclasses 3
values | description stringclasses 3
values | num_tables int64 4 13 | num_rows int64 2.73M 16.9M | num_cols int64 21 174 | num_tasks int64 4 8 | tasks_binary_classification int64 0 3 | tasks_regression int64 0 2 | tasks_multiclass_classification int64 0 8 | start_timestamp stringdate 2000-04-02 00:00:00 2018-01-02 09:22:47 | val_timestamp stringdate 2018-09-01 00:00:00 2022-01-01 00:00:00 | test_timestamp stringdate 2020-01-01 00:00:00 2023-01-01 00:00:00 | size_gb float64 0.15 2.86 | license stringclasses 3
values | source_url stringclasses 3
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
rel-arxiv | Academic citations (arXiv) | arXiv scholarly papers: papers, authors, citations, and subject categories. | 6 | 2,733,846 | 21 | 4 | 1 | 1 | 1 | 2018-01-01 00:00:00 | 2022-01-01 00:00:00 | 2023-01-01 00:00:00 | 0.2232 | CC BY 4.0 | https://github.com/PKUTHM/arxiv-physics |
rel-ratebeer | Beverage reviews | RateBeer reviews: users, beers, brewers, places, and time-stamped ratings. | 13 | 16,874,623 | 174 | 8 | 3 | 2 | 0 | 2000-04-02 00:00:00 | 2018-09-01 00:00:00 | 2020-01-01 00:00:00 | 2.8643 | CC BY-NC-SA 4.0 (as declared by RelBench; no upstream RateBeer license) | https://relbench.stanford.edu |
rel-salt | Enterprise B2B sales (SALT) | SAP SALT sales documents: sales orders, line items, customers, and addresses. | 4 | 4,748,942 | 31 | 8 | 0 | 0 | 8 | 2018-01-02 09:22:47 | 2020-02-01 00:00:00 | 2020-07-01 00:00:00 | 0.1485 | CC BY-NC-SA 4.0 | https://huggingface.co/datasets/SAP/SALT |
RelBench v2 extras
Everything RelBench v2 adds to v1: the v2-only databases with their tasks, and the v2-only tasks on the v1 databases of stanford-star/relbench-v1. A v2-only task on a v1 database is a tasks/ subtree with no db/ beside it; RelBench resolves the database by name.
<dataset>/
manifest.yaml # tables, primary keys, foreign-key graph, val/test timestamps
schema.svg # ER diagram
db/*.parquet # relational tables
tasks/<task>/manifest.yaml # task spec
tasks/<task>/{train,val,test}.parquet
import relbench
relbench.load_dataset("rel-arxiv").load_task("paper-citation") # v2-only database
relbench.load_dataset("rel-f1").load_task("results-position") # v2-only task on a v1 database
License
Each database keeps the license of its upstream source, listed in the license and source_url columns of STATS/databases.parquet. Our contributions (the conversion to this format, schemas, task definitions, splits and statistics) are CDLA-Permissive-2.0.
Citation
@inproceedings{relbenchv2,
title={RelBench v2: A Large-Scale Benchmark and Repository for Relational Data},
author={Gu, Justin and Ranjan, Rishabh and Kanatsoulis, Charilaos and Tang, Haiming and Jurkovic, Martin and Hudovernik, Valter and Znidar, Mark and Chaturvedi, Pranshu and Shroff, Parth and Li, Fengyu and Leskovec, Jure},
booktitle={3rd Workshop on Navigating and Addressing Data Problems for Foundation Models (DATA-FM) at ICLR 2026},
year={2026},
note={arXiv:2602.12606 [cs.LG]},
url={https://arxiv.org/abs/2602.12606}
}
@inproceedings{relbench,
title={RelBench: A Benchmark for Deep Learning on Relational Databases},
author={Robinson, Joshua and Ranjan, Rishabh and Hu, Weihua and Huang, Kexin and Han, Jiaqi and Dobles, Alejandro and Fey, Matthias and Lenssen, Jan Eric and Yuan, Yiwen and Zhang, Zecheng and He, Xinwei and Leskovec, Jure},
booktitle={Advances in Neural Information Processing Systems},
year={2024}
}
- Downloads last month
- 16,547