stanford-star/rt-j
Tabular Classification • 85.6M • Updated • 123 • 2
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
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: string>, targ (... 101 chars omitted)
child 0, item: struct<entity_table: string, kind: string, name: string, splits: list<item: string>, target_col: str (... 89 chars omitted)
child 0, entity_table: string
child 1, kind: string
child 2, name: string
child 3, splits: list<item: string>
child 0, item: string
child 4, target_col: string
child 5, task_type: string
child 6, time_col: string
child 7, remove_columns: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
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
title of product: int64
ltv of item-ltv: int64
brand of product: int64
rating of review: int64
summary of review: int64
review_time of review-rating: int64
churn of item-churn: int64
customer_id of customer: int64
category of product: int64
customer_id of review: int64
rating of review-rating: int64
primary_key of review-rating: int64
timestamp of item-ltv: int64
timestamp of item-churn: int64
product_id of item-churn: int64
timestamp of user-ltv: int64
product_id of review: int64
description of product: int64
timestamp of user-churn: int64
product_id of item-ltv: int64
price of product: int64
review_text of review: int64
ltv of user-ltv: int64
verified of review: int64
customer_id of user-ltv: int64
product_id of product: int64
churn of user-churn: int64
review_time of review: int64
customer_name of customer: int64
customer_id of user-churn: int64
to
{'timestamp of item-ltv': Value('int64'), 'product_id of product': Value('int64'), 'product_id of item-churn': Value('int64'), 'category of product': Value('int64'), 'customer_id of review': Value('int64'), 'rating of review-rating': Value('int64'), 'timestamp of item-churn': Value('int64'), 'price of product': Value('int64'), 'primary_key of review-rating': Value('int64'), 'brand of product': Value('int64'), 'churn of user-churn': Value('int64'), 'verified of review': Value('int64'), 'review_text of review': Value('int64'), 'ltv of item-ltv': Value('int64'), 'review_time of review-rating': Value('int64'), 'churn of item-churn': Value('int64'), 'rating of review': Value('int64'), 'customer_name of customer': Value('int64'), 'timestamp of user-ltv': Value('int64'), 'customer_id of user-ltv': Value('int64'), 'customer_id of user-churn': Value('int64'), 'description of product': Value('int64'), 'title of product': Value('int64'), 'summary of review': Value('int64'), 'timestamp of user-churn': Value('int64'), 'ltv of user-ltv': Value('int64'), 'product_id of review': Value('int64'), 'customer_id of customer': Value('int64'), 'review_time of review': Value('int64'), 'product_id of item-ltv': 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: string>, targ (... 101 chars omitted)
child 0, item: struct<entity_table: string, kind: string, name: string, splits: list<item: string>, target_col: str (... 89 chars omitted)
child 0, entity_table: string
child 1, kind: string
child 2, name: string
child 3, splits: list<item: string>
child 0, item: string
child 4, target_col: string
child 5, task_type: string
child 6, time_col: string
child 7, remove_columns: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
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
title of product: int64
ltv of item-ltv: int64
brand of product: int64
rating of review: int64
summary of review: int64
review_time of review-rating: int64
churn of item-churn: int64
customer_id of customer: int64
category of product: int64
customer_id of review: int64
rating of review-rating: int64
primary_key of review-rating: int64
timestamp of item-ltv: int64
timestamp of item-churn: int64
product_id of item-churn: int64
timestamp of user-ltv: int64
product_id of review: int64
description of product: int64
timestamp of user-churn: int64
product_id of item-ltv: int64
price of product: int64
review_text of review: int64
ltv of user-ltv: int64
verified of review: int64
customer_id of user-ltv: int64
product_id of product: int64
churn of user-churn: int64
review_time of review: int64
customer_name of customer: int64
customer_id of user-churn: int64
to
{'timestamp of item-ltv': Value('int64'), 'product_id of product': Value('int64'), 'product_id of item-churn': Value('int64'), 'category of product': Value('int64'), 'customer_id of review': Value('int64'), 'rating of review-rating': Value('int64'), 'timestamp of item-churn': Value('int64'), 'price of product': Value('int64'), 'primary_key of review-rating': Value('int64'), 'brand of product': Value('int64'), 'churn of user-churn': Value('int64'), 'verified of review': Value('int64'), 'review_text of review': Value('int64'), 'ltv of item-ltv': Value('int64'), 'review_time of review-rating': Value('int64'), 'churn of item-churn': Value('int64'), 'rating of review': Value('int64'), 'customer_name of customer': Value('int64'), 'timestamp of user-ltv': Value('int64'), 'customer_id of user-ltv': Value('int64'), 'customer_id of user-churn': Value('int64'), 'description of product': Value('int64'), 'title of product': Value('int64'), 'summary of review': Value('int64'), 'timestamp of user-churn': Value('int64'), 'ltv of user-ltv': Value('int64'), 'product_id of review': Value('int64'), 'customer_id of customer': Value('int64'), 'review_time of review': Value('int64'), 'product_id of item-ltv': Value('int64')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
RelBench, preprocessed (relbench-preprocessed)
The evaluation and in-loop-validation data for the Relational Transformer: the
seven RelBench databases run through the
rustler preprocessor, in the on-disk format the RT dataloaders read. This is
the pre_dir that reproduces the published RT-J numbers.
- Code: https://github.com/stanford-star/relational-transformer
- Paper: arXiv:2510.06377
- Download / inference:
docs/downloads.md·docs/inference.md
What it is
| Databases | 7 (rel-amazon, rel-avito, rel-event, rel-f1, rel-hm, rel-stack, rel-trial) |
| Tasks | 34 total — 21 forecasting (the RelBench entity benchmark), 13 autocompletion |
| Files | 100 |
| Size | 208,537,743,725 bytes (~194.2 GiB) |
| Text embedder | sentence-transformers/all-MiniLM-L12-v2 (384-d) |
Relation to the raw dataset
Derived from stanford-star/relbench-v1,
the RelBench databases and task tables in RelBench format (parquet +
manifest.yaml), which also carries the regression-target standard deviations
that turn MAE into nMAE. That repository is the one to cite and the one to
re-preprocess from; this one is a build artifact.
File layout
<db>/ # one per database
meta.json table_info.json column_index.json
nodes.rkyv offsets.rkyv p2f_adj.rkyv
text_emb_all-MiniLM-L12-v2.bin
legacy/<db>/ # the same artifacts for the legacy RT-v1 code path
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("relbench", "forecast")
Pass that path as db_task_list.
For this collection the vendored lists are all (34 pairs over the 7
databases), forecast (21 — the RelBench entity benchmark, the default
evaluation list) and autocomplete (13).
Two things rustler writes during preprocessing are not shipped, because
nothing reads them afterwards and shipping them would put verbatim source
content in this repository. text.json, the deduplicated table of source
strings the embedder consumes, is one: 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. legacy/_transformed/, the RelBench-format parquet the
RT-v1 transform emits on its way into rustler, is the other; it is an
intermediate, the legacy path reads legacy/<db>/, and
rt.preprocess.legacy regenerates it from
stanford-star/relbench-v1
whenever it is actually wanted. 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/relbench-preprocessed --repo-type dataset \
--local-dir data/relbench-preprocessed
pixi run python examples/eval.py # pre_dir="data/relbench-preprocessed"
There is no CLI; examples/eval.py is a script that calls rt.eval.main with
every argument spelled out, and the released evaluation values live in it. The
checkpoint is the one thing still resolved from the Hub on demand, so
load_ckpt_path="stanford-star/rt-j" needs no download.
How it was produced
# examples/preprocess.py
one(dataset="stanford-star/relbench-v1/rel-f1", out_dir="data/relbench-preprocessed",
embedder="all-MiniLM-L12-v2", ...) # once per database
There is no CLI: copy
examples/preprocess.py,
edit the call, pixi run python examples/preprocess.py.
Preprocessing. This tree was rebuilt on 2026-10-01 with rustler carrying
PR #4 (rustler: column stats from the train period only, commit 8030aa8),
from the same raw bytes as the first build.
Column statistics
rustler z-scores numeric and datetime cells, and since 8030aa8 the
statistics come from the train period alone rather than from the whole table.
34 of the 50 database tables here are time-indexed and so take narrower
statistics under that change, which is why this tree was rebuilt rather than
left as it was.
Revisions
The RT-J paper's results were produced against revision
1016626ddb30c027b92458bf866903850cc205e1 (2026-08-08), which predates the
rebuild. Hub revisions are immutable, so that pin still resolves to exactly the
data the paper used; pin it to reproduce published numbers:
pixi run hf download stanford-star/relbench-preprocessed --repo-type dataset \
--revision 1016626ddb30c027b92458bf866903850cc205e1 \
--local-dir data/relbench-preprocessed
The rebuilt tree landed at revision
ff29544d3622294d579a05afcdda21b0906283b2 (2026-10-01); later revisions
change only this card.
Licence
CC BY 4.0 — attribution only, commercial use permitted.
The raw collection,
stanford-star/relbench-v1,
is CC BY-SA 4.0 and stays that way: all seven databases are declared
share-alike there, and at least rel-stack inherits that genuinely rather than
by choice, being built from the
Stack Exchange data dumps, whose
user-contributed content is CC BY-SA 4.0.
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 |
Neither the string cell values nor the RelBench-format parquet are here — see File layout for what is left out and why. No string or free-text column can be read back out, the z-scoring statistics are not serialized 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, not the underlying databases and not
the benchmark. To work with the source data, go to
stanford-star/relbench-v1
and follow its terms. Attribution for the databases belongs to RelBench and to
each upstream source: cite RelBench, not only this repository.
Citation
Cite RelBench for the benchmark and the RT paper for this preprocessing:
@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},
}
- Downloads last month
- 929