This repository is the official implementation of the Relational Transformer (RT) architecture for building Relational Foundation Models (RFMs).
| Paper | Venue | Implementation |
|---|---|---|
| Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data | ICLR 2026 | rt-v1 |
| PluRel: Synthetic Data unlocks Scaling Laws for Relational Foundation Models | ICML 2026 | stanford-star/plurel |
| RT-J: Large-Scale Pretraining of Relational Transformers for Context-Efficient Predictions | In progress | main |
Try RT without any local setup: the fully worked Colab notebook (open in Colab) predicts on a general database (which could be your own!) with a released RT-J checkpoint end-to-end. The same flow as plain scripts — with the checkpoint picked straight from the Hugging Face Hub — is in examples/inference.
We use pixi to manage a single, self-contained environment
(Python, PyTorch + CUDA, Rust, and other dependencies).
All commands are run using pixi run to use the environment.
Pixi builds the environment automatically on first use
(check out the docs below).
git clone https://github.com/stanford-star/relational-transformer.git
cd relational-transformer| Docs | Description |
|---|---|
| Downloads | Bulk-download raw data, preprocessed data, and checkpoints from our HuggingFace org |
| Preprocess | Convert RelBench-format databases into RT's on-disk format |
| Inference | Run a trained checkpoint; evaluate, engineer, tune, and ensemble contexts |
| Pretrain | Train RT from scratch, single-GPU to multi-node |
| Baselines | rel2tab tabular baselines through the same eval path |
| Context visualization | Inspect the contexts sampled for each row |