RT-v1

Checkpoints of Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data (ICLR 2026): one .pt file per (database, task). pretrain_<db>_<task>.pt is pretrained with <db> held out, contd-pretrain_* continues pretraining on <db> with the task held out, finetune-from-* is fine-tuned on the task. 12 blocks, d_model 256, 8 heads, d_ff 1024; text embedded with all-MiniLM-L12-v2.

Superseded by stanford-star/rt-j. The original implementation is the rt-v1 branch; main evaluates these checkpoints via examples/eval/legacy.py against the legacy/ subdirectory of relbench-preprocessed.

License

Apache-2.0.

Citation

@inproceedings{ranjan2026relational,
    title={Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data},
    author={Rishabh Ranjan and Valter Hudovernik and Mark Znidar and Charilaos Kanatsoulis and Roshan Upendra and Mahmoud Mohammadi and Joe Meyer and Tom Palczewski and Carlos Guestrin and Jure Leskovec},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026}
}