Datasets:

stanford-star
/

relbench-raw

Size:

10B<n<100B

Tags:

relational-deep-learning

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

The dataset viewer is not available for this subset.

Cannot get the split names for the config 'default' of the dataset.

Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/parquet/parquet.py", line 127, in _split_generators
                  self.info.features = datasets.Features.from_arrow_schema(pq.read_schema(f))
                                                                           ~~~~~~~~~~~~~~^^^
                File "/usr/local/lib/python3.14/site-packages/pyarrow/parquet/core.py", line 2393, in read_schema
                  file = ParquetFile(
                      where, memory_map=memory_map,
                      decryption_properties=decryption_properties)
                File "/usr/local/lib/python3.14/site-packages/pyarrow/parquet/core.py", line 328, in __init__
                  self.reader.open(
                  ~~~~~~~~~~~~~~~~^
                      source, use_memory_map=memory_map,
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ...<8 lines>...
                      arrow_extensions_enabled=arrow_extensions_enabled,
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "pyarrow/_parquet.pyx", line 1656, in pyarrow._parquet.ParquetReader.open
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
              pyarrow.lib.ArrowInvalid: Parquet magic bytes not found in footer. Either the file is corrupted or this is not a parquet file.
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

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.

RelBench raw sources (`relbench-raw`)

A mirror of the upstream raw sources behind RelBench databases and the TGB datasets, kept so RelBench-format databases can be rebuilt from scratch without depending on third-party hosts staying up. These are the inputs to the database construction step, not to RT.

Code: https://github.com/stanford-star/relational-transformer
RelBench: https://relbench.stanford.edu
TGB: https://tgb.complexdatalab.com

What it is


Files	23 (22 zip archives + `.gitattributes`)
Size	~30.6 GiB

rel-arxiv/db.zip                       114.0 MB
rel-avito/rel-avito-raw-100k.zip       518.7 MB
rel-f1/relbench-f1-raw.zip               6.2 MB
rel-ratebeer/db.zip                      1.9 GB
rel-salt/db.zip                         35.3 MB
rel-stack/relbench-forum-raw.zip       700.4 MB
rel-trial/relbench-trial.zip             1.1 GB
tgb/tgbl-coin/db.zip                     1.9 GB
tgb/tgbl-comment/db.zip                  3.7 GB
tgb/tgbl-flight/db.zip                   2.6 GB
tgb/tgbl-review/db.zip                   1.3 GB
tgb/tgbl-review-v2/db.zip                1.3 GB
tgb/tgbl-wiki/db.zip                   383.5 MB
tgb/tgbl-wiki-v2/db.zip                384.4 MB
tgb/tgbn-genre/db.zip                  133.8 MB
tgb/tgbn-reddit/db.zip                 304.5 MB
tgb/tgbn-token/db.zip                  308.2 MB
tgb/tgbn-trade/db.zip                    6.5 MB
tgb/thgl-forum/db.zip                    6.5 GB
tgb/thgl-github/db.zip                   1.6 GB
tgb/thgl-myket/db.zip                    4.4 GB
tgb/thgl-software/db.zip                 3.7 GB

Relation to the other repositories

Three layers, in order:

layer	repository	what it holds
raw upstream	this repository	the original vendor dumps, as zips
RelBench format	`stanford-star/relbench-v1`, `stanford-star/tgb`	parquet tables + `manifest.yaml`
RT format	`stanford-star/relbench-preprocessed`	`rustler` artifacts + text embeddings

You want this repository only if you are rebuilding a RelBench-format database from its original source. To train or evaluate RT, go straight to the preprocessed repository; to re-run RT's preprocessing, use the RelBench-format repository.

How it was produced

Each archive is a byte mirror of an upstream download, uploaded with huggingface_hub and committed one dataset at a time (the commit titles record the source, e.g. "Mirror the rel-ratebeer raw source"). Nothing in this repository is produced by this project's code, so there is no preprocessing commit for it — and conversely, it is unaffected by the rustler normalization fix (PR #4, commit 8030aa8) that changes the preprocessed repositories.

Revisions

The RT-J paper's results do not read this repository directly, but the RelBench-format databases they do read were built from revision f1d7228af23a22b9ece756fe5dda4dda79b711dc (2026-08-25, the head at the time). Pin it:

pixi run hf download stanford-star/relbench-raw --repo-type dataset \
  --revision f1d7228af23a22b9ece756fe5dda4dda79b711dc

Licence

Read this before redistributing anything here. The cc-by-4.0 in the front matter covers this repository's packaging only — the selection, naming and archiving of the mirrored files. It is not a licence for the data inside the archives.

Every archive retains the licence and terms of its upstream source. This is a byte mirror of third-party dumps: nothing in it is this project's work, and this repository does not and cannot relicense it. RelBench's underlying sources and the TGB datasets each carry their own terms, some of them share-alike or otherwise restrictive, and several require attribution to the original publisher rather than to us.

So:

To use an archive, follow its upstream licence.
To redistribute an archive, check its upstream licence first. The packaging licence above does not grant you that right.
Per-source terms for the RelBench-format databases are catalogued in STATS/databases.parquet on stanford-star/relbench-v1; TGB's are documented at https://tgb.complexdatalab.com.

RelBench raw sources (relbench-raw)