PluRel, preprocessed (plurel-preprocessed)

Phase 1 of RT-J's pretraining, and the pretraining corpus of RT-PluRel: the synthetic relational databases of stanford-star/plurel run through the rustler preprocessor, in the on-disk format the Relational Transformer dataloaders read.

• Code: https://github.com/stanford-star/relational-transformer
• Papers: Relational Transformer (arXiv:2510.06377) · PluRel (arXiv:2602.04029)

What it is

Databases	2,000 synthetic (plurel-3000 … plurel-4999)
Files	16,005
Size	43,218,968,510 bytes (~40.2 GiB)
Text embedder	sentence-transformers/all-MiniLM-L12-v2 (384-d)

Per-file-kind totals: nodes.rkyv ~35.2 GiB, p2f_adj.rkyv ~4.1 GiB, offsets.rkyv ~0.8 GiB, text_emb_all-MiniLM-L12-v2.bin ~0.1 GiB (synthetic databases carry little text).

The curated pretraining mixture is db-task-lists/rt-plurel-train.json: 86,211 (db, task) pairs over 1,900 of the 2,000 databases — the filtered subset RT-J's phase 1 and RT-PluRel actually train on. The remaining 100 databases are present but excluded by the filter.

Relation to the raw dataset

Derived from stanford-star/plurel — synthetic databases in RelBench format (parquet + manifest.yaml), generated by PluRel. That repository is the one to generate from or cite; this one is a build artifact and is not a substitute.

A 2026-09-12 commit dropped the older generation of databases, leaving only plurel-3000 … plurel-4999; the same round regenerated the task lists with autocomplete task manifests and the filtered rt-plurel-train.json. An earlier revision is a different collection.

File layout

db-task-lists/
  all.json
  forecast.json
  autocomplete.json
  rt-plurel-train.json   # the curated 86,211-task / 1,900-db training mixture
plurel-<n>/              # one per synthetic database, n = 3000..4999
  meta.json  table_info.json  column_index.json
  nodes.rkyv  offsets.rkyv  p2f_adj.rkyv
  text_emb_all-MiniLM-L12-v2.bin
  text.json

pre_dir is always a local directory; download with hf download --local-dir, nothing is fetched on demand.

How it was produced

# examples/preprocess.py, preprocess_a_collection()
many(repo="stanford-star/plurel", out_dir="data/plurel-preprocessed",
     shard=0, num_shards=1, skip_existing=True, embedder="all-MiniLM-L12-v2", ...)

There is no CLI: copy examples/preprocess.py, edit the call, pixi run python examples/preprocess.py. The curated task list is built by expts/preprocess/plurel_filtered_list.py, which writes db-task-lists/rt-plurel-train.json.

Preprocessing commit. The 2026-09-12 content was written by rustler at repository commit 970167c (2026-09-09, rustler: no boolean sem type in sampler output; bools are z-scored numbers end to end), the last preprocessing-code commit before that rebuild.

This is pre-fix, leaky normalization. rustler PR #4 (rustler: column stats from the train period only, commit 8030aa8, merged 2026-10-01) restricted z-scoring statistics to the train period; before it, column statistics and the global datetime statistics were computed over the validation and test rows too. The on-disk format is unchanged, so this dataset keeps the old, leaky normalization until it is regenerated. RT-J phase 1 and RT-PluRel were both trained on this pre-#4 data. The fix engages here because PluRel manifests carry a real val_timestamp; it does not engage on the-join-preprocessed, whose manifests have val_timestamp: null. For a measured bound on how far the fix moves downstream numbers, see the relbench-preprocessed card: mean −0.045 AUROC and −0.035 nMAE over seven paired tasks, mixed in sign, one context seed. That bounds the magnitude; it does not correct any published number.

Revisions

The RT-J paper's results were produced against revision 9d70172425b44053f1270b19094ba6dcf7d66464 (2026-09-12). Pin it:

pixi run hf download stanford-star/plurel-preprocessed --repo-type dataset \
  --revision 9d70172425b44053f1270b19094ba6dcf7d66464 \
  --local-dir data/plurel-preprocessed

Licence

CC BY 4.0 — attribution only, commercial use permitted.

Note the asymmetry with the upstream collection, which is deliberate. The databases here are entirely synthetic: they were generated by PluRel, contain no third-party data, and carry no inherited terms. stanford-star holds the rights in both the upstream collection and this derived artifact, so this repository can be offered under the more permissive CC BY 4.0 even though stanford-star/plurel is published as CC BY-SA 4.0. The upstream collection keeps its own CC BY-SA 4.0 terms when you redistribute it; CC BY 4.0 here applies to this preprocessed build.

This reasoning does not carry over to the preprocessed builds of real-world data (the-join-preprocessed, relbench-preprocessed), whose source databases are third-party and share-alike.

Citation

@article{ranjan2026rtj,
  title     = {RT-J: Large-Scale Pretraining of Relational Transformers for
               Context-Efficient Predictions},
  author    = {Ranjan, Rishabh and Kothapalli, Vignesh and Agarwal,
               Harshvardhan and Kanatsoulis, Charilaos and Upendra, Roshan and
               Palczewski, Tom and Guestrin, Carlos and Leskovec, Jure},
  year      = {2026},
}