Datasets:

stanford-star
/

the-join-preprocessed

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
files: struct<column_index: string, nodes: string, offsets: string, p2f_adj: string, table_info: string>
  child 0, column_index: string
  child 1, nodes: string
  child 2, offsets: string
  child 3, p2f_adj: string
  child 4, table_info: string
format_version: int64
name: string
num_db_tables: int64
num_nodes: int64
num_task_tables: int64
num_text_strings: int64
source: string
tasks: list<item: struct<entity_table: string, kind: string, name: string, splits: list<item: null>, target (... 49 chars omitted)
  child 0, item: struct<entity_table: string, kind: string, name: string, splits: list<item: null>, target_col: strin (... 37 chars omitted)
      child 0, entity_table: string
      child 1, kind: string
      child 2, name: string
      child 3, splits: list<item: null>
          child 0, item: null
      child 4, target_col: string
      child 5, task_type: string
      child 6, time_col: null
text_embeddings: struct<all-MiniLM-L12-v2: struct<file: string, d_text: int64>>
  child 0, all-MiniLM-L12-v2: struct<file: string, d_text: int64>
      child 0, file: string
      child 1, d_text: int64
name of train: int64
annual_interchanges of station: int64
train_id of train: int64
name of station: int64
number_of_platforms of station: int64
identifier of train_station: int64
station_id of train_station: int64
station_id of station: int64
time of train: int64
annual_entry_exit of station: int64
train_id of train_station: int64
total_passengers of station: int64
location of station: int64
main_services of station: int64
to
{'station_id of train_station': Value('int64'), 'main_services of station': Value('int64'), 'time of train': Value('int64'), 'name of train': Value('int64'), 'total_passengers of station': Value('int64'), 'station_id of station': Value('int64'), 'train_id of train': Value('int64'), 'name of station': Value('int64'), 'train_id of train_station': Value('int64'), 'location of station': Value('int64'), 'identifier of train_station': Value('int64'), 'annual_entry_exit of station': Value('int64'), 'annual_interchanges of station': Value('int64'), 'number_of_platforms of station': Value('int64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              files: struct<column_index: string, nodes: string, offsets: string, p2f_adj: string, table_info: string>
                child 0, column_index: string
                child 1, nodes: string
                child 2, offsets: string
                child 3, p2f_adj: string
                child 4, table_info: string
              format_version: int64
              name: string
              num_db_tables: int64
              num_nodes: int64
              num_task_tables: int64
              num_text_strings: int64
              source: string
              tasks: list<item: struct<entity_table: string, kind: string, name: string, splits: list<item: null>, target (... 49 chars omitted)
                child 0, item: struct<entity_table: string, kind: string, name: string, splits: list<item: null>, target_col: strin (... 37 chars omitted)
                    child 0, entity_table: string
                    child 1, kind: string
                    child 2, name: string
                    child 3, splits: list<item: null>
                        child 0, item: null
                    child 4, target_col: string
                    child 5, task_type: string
                    child 6, time_col: null
              text_embeddings: struct<all-MiniLM-L12-v2: struct<file: string, d_text: int64>>
                child 0, all-MiniLM-L12-v2: struct<file: string, d_text: int64>
                    child 0, file: string
                    child 1, d_text: int64
              name of train: int64
              annual_interchanges of station: int64
              train_id of train: int64
              name of station: int64
              number_of_platforms of station: int64
              identifier of train_station: int64
              station_id of train_station: int64
              station_id of station: int64
              time of train: int64
              annual_entry_exit of station: int64
              train_id of train_station: int64
              total_passengers of station: int64
              location of station: int64
              main_services of station: int64
              to
              {'station_id of train_station': Value('int64'), 'main_services of station': Value('int64'), 'time of train': Value('int64'), 'name of train': Value('int64'), 'total_passengers of station': Value('int64'), 'station_id of station': Value('int64'), 'train_id of train': Value('int64'), 'name of station': Value('int64'), 'train_id of train_station': Value('int64'), 'location of station': Value('int64'), 'identifier of train_station': Value('int64'), 'annual_entry_exit of station': Value('int64'), 'annual_interchanges of station': Value('int64'), 'number_of_platforms of station': Value('int64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

The Join, preprocessed (`the-join-preprocessed`)

The pretraining corpus of RT-J, in the on-disk format the Relational Transformer dataloaders read. It is stanford-star/the-join — real-world multi-table databases in RelBench format — run through the rustler preprocessor, plus frozen text-column embeddings.

Code: https://github.com/stanford-star/relational-transformer
Paper: arXiv:2510.06377
How to download a subset: docs/downloads.md
How it is produced: docs/preprocess.md

What it is


Databases	523 (those under the 5 GB per-database cutoff, `all_5gb_cutoff`)
Tasks	13,243 (db, task) pairs over those 523 databases
Files	3,667
Size	275,371,755,983 bytes (~256.5 GiB)
Text embedder	`sentence-transformers/all-MiniLM-L12-v2` (384-d)

Per-file-kind totals, which is what a partial download is planned against:

file	total
`nodes.rkyv`	~195.1 GiB
`text_emb_all-MiniLM-L12-v2.bin`	~29.8 GiB
`p2f_adj.rkyv`	~26.1 GiB
`offsets.rkyv`	~5.3 GiB

Relation to the raw dataset

stanford-star/the-join is the raw side: parquet tables plus a manifest.yaml of relational metadata, per database, in RelBench format. This repository is derived from it and is not a substitute — the raw repository is what you need to re-run preprocessing, and it carries the per-database licence attribution.

The RT-J paper describes the collection at 639 databases, which is the untrimmed raw side. This preprocessed repository was trimmed on 2026-09-12 to the 523 all_5gb_cutoff databases (116 dropped), and those 523 are the set every published RT-J number was pretrained on. A reader who pinned an earlier revision gets a different, larger dataset; see Revisions below.

File layout

One directory per database:

<db>/
  meta.json                            # relational + task metadata rustler emits
  table_info.json                      # table names, row counts
  column_index.json                    # column -> index, semantic type
  nodes.rkyv                           # rkyv-serialized cell values (the bulk)
  offsets.rkyv                         # row offsets into nodes
  p2f_adj.rkyv                         # primary->foreign key adjacency
  text_emb_all-MiniLM-L12-v2.bin       # frozen text-column embeddings

The task lists are not in this repository. The curated (db, task) mixtures are vendored in the rt package instead, so a list can no longer drift from the data or the code that reads it:

from rt.data import get_mixture_path, list_mixtures

list_mixtures()                                    # every (collection, name)
path = get_mixture_path("the-join", "rt-j")

Pass that path as db_task_list.

For this collection the vendored lists are: rt-j and all (both 13,243 pairs over the 523 databases — they are the same list, and rt-j is RT-J's phase-2 pretraining mixture), autocomplete (9,145 over 519) and forecast (4,098 over 153).

There is no text.json. rustler writes one during preprocessing — the deduplicated table of source strings that the embedder consumes — and it is not shipped, because nothing reads it after the embeddings exist: a string cell in nodes.rkyv is an index, and training and inference gather the embedding at that index straight out of text_emb_*.bin. Removing it is also what makes the licence below true. The cost to you: a downloaded tree can no longer be re-embedded with a different text model, because the strings it would need are not here. Using a different text embedder means re-running preprocessing from the raw repository (see docs/preprocess.md).

How to load it

pre_dir is always a local directory — nothing is fetched on demand. Download it, then point a run at the path:

pixi run hf download stanford-star/the-join-preprocessed --repo-type dataset \
  --local-dir data/the-join-preprocessed
pixi run python examples/train.py        # pre_dir="data/the-join-preprocessed"

There is no CLI; examples/train.py is a script that calls rt.train.main with every argument spelled out, and the released pretraining values live in it. Its db_task_list comes from rt.data.get_mixture_path("the-join", "rt-j"). docs/downloads.md has the --include patterns that skip the embedders you are not using, and the revision pins below.

How it was produced

# examples/preprocess.py, preprocess_a_collection()
many(repo="stanford-star/the-join", out_dir="data/the-join-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. shard=i, num_shards=N splits the collection across a job array.

Preprocessing commit. The content of this repository was written by rustler at repository commit dc56007 (2026-08-05, the last preprocessing-code commit before the upload on 2026-08-05/07); the 2026-09-12 commits are a trim and a task-list regeneration, not a re-preprocessing.

Column statistics

rustler z-scores numeric and datetime cells. Since PR #4 (rustler: column stats from the train period only, commit 8030aa8) the statistics come from the train period alone. That change leaves this collection byte-identical: every manifest in stanford-star/the-join has val_timestamp: null and the collection has no validation or test splits, so neither half of the fix engages. The databases here were therefore not rebuilt — the artifacts at the head are byte-identical to the ones the RT-J paper used, and the head has moved only for metadata (the card, and dropping text.json). Unlike relbench-preprocessed and plurel-preprocessed, which the fix does reach and which were rebuilt on 2026-10-01.

PR #3 (rustler: draw BFS children without rejection sampling, fly.rs) changed the context sampler, not this data.

Revisions

The RT-J paper's results were produced against revision ba5574ba659f0ef8b592793ed2ac9cb78dc87100 (2026-09-12, the completed trim). Pin it:

pixi run hf download stanford-star/the-join-preprocessed --repo-type dataset \
  --revision ba5574ba659f0ef8b592793ed2ac9cb78dc87100 \
  --local-dir data/the-join-preprocessed

Licence

CC BY 4.0 — attribution only, commercial use permitted.

The raw collection, stanford-star/the-join, is CC BY-SA 4.0 and stays that way: it is an aggregate of 639 third-party databases, about half of them share-alike upstream (Spider 1.0 143, BIRD 58, Stack Exchange dumps 40, Wikipedia-derived 5, ODbL 4, CC BY-SA 3.0 3, GPL 2.0 1), and that obligation is inherited rather than chosen. Per-database attribution — the license and source_url columns for all 639 — is in STATS/databases.parquet there.

This repository can be more permissive because it contains no verbatim upstream content. Per database it holds, and only holds:


`text_emb_*.bin`	one 384-d `bf16` MiniLM embedding per distinct source string
`nodes.rkyv`	numeric and datetime cells z-scored, text cells as an embedding index, raw row timestamps
`offsets.rkyv`, `p2f_adj.rkyv`	row offsets and the foreign-key graph
`meta.json`, `table_info.json`, `column_index.json`	table and column names, row counts, semantic types

The string cell values themselves are not here: text.json is a preprocessing intermediate and is not shipped, so no string or free-text column can be read back out. The z-scoring statistics are not serialized either, so numeric scale is unrecoverable, and primary-key columns are dropped. What remains is a lossy derived representation plus schema metadata — the same basis on which the released checkpoints are CC BY 4.0.

So CC BY 4.0 covers these artifacts. It is not a licence for the underlying databases, and it is not a route around their terms:

To work with the source data, go to stanford-star/the-join and follow each database's own licence.
Cite the RT-J paper for this artifact, and the upstream sources for the data behind it. Several require attribution to the original publisher.
Twelve of the 523 databases carry upstream terms stricter than share-alike — join-yoochoose (CC BY-NC-ND 4.0), join-apple-podcasts and join-atp-tennis (CC BY-NC-SA 4.0), join-imdb-full, join-imdb-ijs, join-imdb-small, join-imsa, join-indycar, join-nascar-cup, join-nascar-trk (non-commercial, several also no-redistribution), and join-gbif-biodiversity, join-gbif-species (mixed per occurrence record). Those terms bind the source data, not the derived vectors here, but if you intend to reconstruct or redistribute anything resembling the originals, start from the raw repository and read them.

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},
}