Stanford Tabular and Relational Project
Advancing foundation models for structured data, from a single table to the many linked tables of a relational database.
The Initiative
Foundation models have transformed natural language, vision, and code, and a recent wave of tabular foundation models is extending this paradigm to individual tables. We develop foundation models for structured data across this spectrum, from a single table to the many interconnected tables of a relational database, where each new schema has traditionally required a model trained from scratch.
We release the synthetic and real-world datasets used to train these models, and make all papers, code, and weights openly available.
Learn morePublications
Selected publications from the group. Papers, code, datasets, and model weights are openly available.
Open Resources
Everything is open. Models and datasets live on Hugging Face, the source code on GitHub, and the benchmark and leaderboard on RelBench.
Datasets
| Dataset | Type | |
|---|---|---|
| relbench | benchmark | open ↗ |
| relbench-v2-extra | benchmark | open ↗ |
| plurel | synthetic | open ↗ |
| the-join | corpus | open ↗ |
| redelex | corpus | open ↗ |
| tgb | temporal | open ↗ |
| dbinfer | tasks | open ↗ |
Dataset descriptions summarize their role in the initiative. See each card on Hugging Face for authoritative details.
News
Recent paper releases, software, and announcements, with related artifacts linked inline.
The synthetic-data and scaling-laws work for relational foundation models is presented at the International Conference on Machine Learning.
RT-J pretrains a Relational Transformer on THE JOIN (6,255 forecasting tasks across 650 real-world databases) to make state-of-the-art predictions from only hundreds of in-context labels, matching strong in-context-learning pipelines with 23–32× fewer examples. The corpus and model weights are released on Hugging Face.
Accepted to ICML 2026. The paper and dataset show that synthetically generated relational databases exhibit power-law scaling and transfer to real-world RelBench tasks; the synthetic databases and code are published on Hugging Face and GitHub. arXiv ↗
A large-scale benchmark and repository for relational data, expanding the standardized datasets and tasks tracked on the RelBench leaderboard. arXiv ↗
The Relational Transformer and the Relational Graph Transformer are accepted to the International Conference on Learning Representations.
A 22M-parameter model attains 93% of fully-supervised AUROC without fine-tuning, outperforming a 27B-parameter language model. Code and weights are openly available. arXiv ↗
RelGNN (ICML 2025) and Rel-LLM (ACL 2025) advance relational architectures, and a survey of relational deep learning appears at KDD 2025.
An open benchmark of seven databases and thirty predictive tasks for deep learning on relational data, presented in the Datasets and Benchmarks track. relbench.stanford.edu ↗
The founding paper framing relational databases as graphs to enable end-to-end learning without manual feature engineering. PMLR ↗
Talks
Recorded talks on the group's research, hosted on YouTube.
PluRel: Synthetic Data unlocks Scaling Laws for Relational Foundation Models
Vignesh Kothapalli · Temporal Graph Learning Reading Group
Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data
Rishabh Ranjan · Temporal Graph Learning Reading Group
Relational Foundation Models
Jure Leskovec · ACM RecSys 2025 keynote
Large Language Models are Good Relational Learners
Fang Wu & Vijay Prakash Dwivedi · Temporal Graph Learning Reading Group
RelBench: A Benchmark for Deep Learning on Relational Databases
Rishabh Ranjan · Temporal Graph Learning Reading Group
People
