RT-PluRel (rt-plurel)

RT-PluRel is a Relational Transformer checkpoint pair for in-context / few-shot entity prediction over multi-table relational databases (no per-task gradient training). Same architecture and file layout as stanford-star/rt-j — drop-in replacement.

Repository layout note (2026-08). This repository now hosts two generations of RT-PluRel checkpoints. The current checkpoint pair (formerly published as stanford-star/rt-p, which this repository replaces) lives at the repository root, described below. The original checkpoints of the PluRel paper (arXiv:2602.04029), previously at this repository's root, are preserved unchanged under paper/ — see paper/README.md.

Variant	Folder	Task
Classifier	classification/	binary entity classification
Regressor	regression/	entity regression

Architecture: ~85.6M params · bfloat16 · 12 blocks, d_model 512, 8 heads, d_ff 2048 · text columns embedded with all-MiniLM-L12-v2 (d_text 384). Each folder contains model.safetensors (weights) and config.json (dims + text-embedding model).

Usage

from rt.model import RelationalTransformer

model = RelationalTransformer.from_pretrained("stanford-star/rt-plurel/classification")
# or: RelationalTransformer.from_pretrained("stanford-star/rt-plurel", subfolder="regression")

pixi run eval --model.load-ckpt-path stanford-star/rt-plurel/classification ...

paper/: original PluRel-paper checkpoints

The paper/ subdirectory preserves the earlier RT-PluRel release: .pt checkpoints (12 blocks, d_model 256, d_ff 1024) pretrained on synthetic relational databases generated by PluRel, plus the continued-pretraining and fine-tuned RelBench leaderboard checkpoints. Details, protocols, and download snippets: paper/README.md. Paths from before the reorganization gain a paper/ prefix, e.g.

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("stanford-star/rt-plurel", "paper/synthetic-pretrain_rdb_1024_size_4b.pt")

License

The root checkpoint pair (classification/, regression/) is released under CC BY-NC-SA 4.0 (see LICENSE). The files under paper/ retain their original MIT license.

Related

• Training data: stanford-star/the-join
• Evaluation: RelBench
• Reference model: stanford-star/rt-j
• Papers: Relational Transformer (arXiv:2510.06377) · PluRel (arXiv:2602.04029)