---

license: cc-by-4.0 tags:

• relbench
• relational-deep-learning
• temporal-graph pretty_name: TGB in RelBench format configs:
• config_name: databases data_files:
  • split: eval path: STATS/databases.parquet
• config_name: tasks data_files:
  • split: eval path: STATS/tasks.parquet---

TGB in RelBench format

The Temporal Graph Benchmark (TGB) datasets in RelBench format. Labels, splits and official negative samples are TGB's, served as-is (kind: external, evaluator: tgb). Each dataset ships negatives/ and, for heterogeneous graphs, mappings/, which the TGB protocol needs.

family	datasets	task
tgbl-* link prediction	tgbl-wiki, tgbl-wiki-v2, tgbl-review, tgbl-review-v2, tgbl-coin, tgbl-comment, tgbl-flight	src-dst-mrr
thgl-* heterogeneous link prediction	thgl-software, thgl-forum, thgl-github, thgl-myket	edge-type-<k>-mrr
tgbn-* node property prediction	tgbn-trade	node-label-ndcg

import relbench
ds = relbench.load_dataset("stanford-star/tgb/tgbl-wiki")
task = relbench.load_task("stanford-star/tgb/tgbl-wiki", "src-dst-mrr")

Citation

Cite the TGB papers (Huang et al., 2023 and the THGB extension; see https://tgb.complexdatalab.com/) and follow the license of each source dataset. For the RelBench-format port:

@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}
}