RT-PluRel

Relational Transformer checkpoint pretrained only on synthetic databases from PluRel. Same architecture and file layout as stanford-star/rt-j, which is warm-started from it and is the stronger model. #1 on the combined leaderboard of RelArena-alpha at submission time.

85M parameters, 12 blocks, d_model 512, 8 heads, d_ff 2048; text columns embedded with all-MiniLM-L12-v2. Files: model.safetensors, config.json. The paper/ directory holds the original checkpoints of the PluRel paper (.pt, d_model 256), described in paper/README.md.

Code and recipes: https://github.com/stanford-star/relational-transformer. Paper: PluRel (ICML 2026).

from rt.model import load_rt_model
model = load_rt_model("stanford-star/rt-plurel")

License

Apache-2.0, including the files under paper/.

Citation

@inproceedings{kothapalli2026plurel,
    title={{PluRel}: Synthetic Data unlocks Scaling Laws for Relational Foundation Models},
    author={Vignesh Kothapalli and Rishabh Ranjan and Valter Hudovernik and Vijay Prakash Dwivedi and Johannes Hoffart and Carlos Guestrin and Jure Leskovec},
    booktitle={Forty-third International Conference on Machine Learning},
    year={2026}
}

@inproceedings{ranjan2026rtj,
    title={{RT-J}: Large-Scale Pretraining of Relational Transformers for Context-Efficient Predictions},
    author={Rishabh Ranjan and Vignesh Kothapalli and Harshvardhan Agarwal and Charilaos Kanatsoulis and Roshan Upendra and Tom Palczewski and Carlos Guestrin and Jure Leskovec},
    booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
    year={2026}
}