BeyondArena Datasets

Datasets from BeyondArena, a unified, holistic benchmark for tabular data that supports diverse task types (IID, temporal, grouped), across sample size and feature dimensionality scales, with diverse feature types (with text, with high cardinality) from a broad range of disciplines.

We introduce BeyondArena and its datasets in Beyond IID: How General Are Tabular Foundation Models, Really?.

@misc{purucker2026iidgeneraltabularfoundation,
      title={Beyond IID: How General Are Tabular Foundation Models, Really?}, 
      author={Lennart Purucker and Andrej Tschalzev and Nick Erickson and Gioia Blayer and David Holzmüller and Alan Arazi and Alexander Pfefferle and Mustafa Tajjar and Gaël Varoquaux and Frank Hutter},
      year={2026},
      eprint={2606.30410},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2606.30410}, 
}

More details:

• Project page and leaderboard: http://tabarena.ai/
• Code / Benchmark repository: https://tabarena.ai/code

Quickstart

We recommend using the datasets via Data Foundry, which resolves a curated container (table + dtypes + task metadata + outer CV splits) by name and caches it locally:

pip install data-foundry

from data_foundry.collections import BEYOND_ARENA

container = BEYOND_ARENA.get_dataset("airfoil_self_noise")
print(container.describe())                  # full identity + dtypes + task + splits
print(container.dataset.shape)               # the actual DataFrame
print(container.task_metadata.split_regime)  # "iid", "temporal_non_iid", or "grouped_non_iid"

df = container.dataset
target = container.task_metadata.target_column_name

for repeat_id, folds in container.experiment_metadata.splits.items():
    for fold_id, (train_idx, test_idx) in folds.items():
        X_train, y_train = df.iloc[train_idx].drop(columns=target), df.iloc[train_idx][target]
        X_test,  y_test  = df.iloc[test_idx].drop(columns=target),  df.iloc[test_idx][target]
        # ... fit, evaluate ...

To pre-download the entire collection in a single network round-trip:

from data_foundry.collections import BEYOND_ARENA

BEYOND_ARENA.prefetch()                          # warms the cache once
for container in BEYOND_ARENA.iter_containers(): # now hits disk only
    print(container.dataset_metadata.unique_name, container.dataset.shape)

See Data Foundry's examples for a full benchmarking walkthrough, the three split regimes (IID / temporal / grouped), and the curation flow.

Datasets

BeyondArena comes with 142 datasets. BeyondArena covers tabular classification and regression tasks. And the following types of datasets:

• IID tabular data
• Non-IID temporal tabular data
• Non-IID grouped tabular data
• IID and non-IID tabular data with text features
• Tabular data with high-cardinality categoricals

Dataset Selection Overview

We build on top of the dataset curation protocol of TabArena-v0.1 (https://arxiv.org/abs/2506.16791) and curate 142 tiny to large-sized, tabular IID and non-IID tasks. For details, see the paper.

Dataset Dashboard

We curated a diverse set of datasets. We share the dataset sizes (w.r.t. rows, columns, and cells), their age distribution, the distribution of feature types per dataset, and the share of datasets from a specific problem type, task type, dataset source, or application domain.

Per-Dataset Index

Per-dataset metadata for the BeyondArena benchmark, sorted by number of rows (N).

Dataset Structure

The release ships as a flat bundle of 142 datasets. Each dataset lives in its own top-level directory named by unique_name, with a UUID-named version subdirectory holding all artifacts. Two layout variants exist:

<dataset_name>/<uuid>/...                 # default (132 datasets)
<dataset_name>/versions/<uuid>/...        # versioned wrapper (10 large non-IID datasets)

Each directory contains six core files, plus an optional tabarena_text_cache.parquet for the 16 datasets with text features:

<uuid>/
├── dataset.parquet                                       # the table (rows × columns)
├── dtypes.json                                           # column name → pandas dtype
├── container_metadata.json                               # uuid + sha256 checksum
├── dataset_metadata.dataset-mold-v1.json                 # provenance & curation notes
├── task_metadata.predictive-ml-task-mold-v1.json         # target, problem type, metric, split keys
├── experiment_metadata.predictive-ml-splits-mold-v1.json # CV fold indices
└── tabarena_text_cache.parquet                           # (optional) precomputed sentence embeddings keyed by text

For details on files and the metadata structure, checkout DataFoundry!

Text embedding cache (tabarena_text_cache.parquet)

Shipped for the 16 text-bearing datasets listed below. The file is a pandas DataFrame written via SemanticTextFeatureGenerator.save_embedding_cache (see TabArena's text_feature_generators.py):

• Index — a string column named text containing every unique text value observed across all text columns of dataset.parquet.
• Columns — 0, 1, …, D-1, holding the precomputed sentence embedding for each text value (default model: 32-dim embeddings).

Reload with:

import pandas as pd
df = pd.read_parquet("<dataset_name>/<uuid>/tabarena_text_cache.parquet")
cache = dict(zip(df.index, df.to_numpy()))   # {text: np.ndarray}

This lets you skip the embedding step at fit time. Datasets with a tabarena_text_cache.parquet: coffee_rating_prediction, consumer_complaints (1m variant), california_house_prices_2020, drug_induced_autoimmunity_prediction, emscad, immoscout_german_house_prices, kickstarter, labour_inspection_compliance, lending_club (1m variant), mercari_price_suggestion (1m variant), mutual_funds_india, pva_revenue_prediction_kddcup98, regensburg_pediatric_appendicitis, sf_permit_time, wids_diabetes_mellitus, wine_world_cost.

Loading a single dataset directly

Each per-dataset config in this card's frontmatter routes only dataset.parquet, which is enough to get the table but not the sibling metadata files (dtypes.json, task_metadata.*, experiment_metadata.* with the CV folds, dataset_metadata.*, container_metadata.json). Because the benchmark protocol depends on those files, the recommended path is to download the whole dataset folder with huggingface_hub:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="TabArena/BeyondArena",
    repo_type="dataset",
    allow_patterns=["churn/**"],                # one or more <dataset_name>/** globs
)
# local_dir/<dataset_name>/<uuid>/ now contains all six files for that dataset.

For the 10 datasets that use the versions/ wrapper (see Dataset Structure), the layout is <dataset_name>/versions/<uuid>/... — the <dataset_name>/** glob already covers both layouts.

If you only need the table (no folds, no metadata), the datasets library shortcut works:

from datasets import load_dataset

ds = load_dataset("<org>/BeyondArena", name="churn")   # any per-dataset config_name

Downloading the full bundle

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="<org>/BeyondArena",
    repo_type="dataset",
)

Licensing

This collection is released under the terms in LICENSE (copyright-at-original-authors). Individual datasets retain their original licenses; see each dataset metadata for their source-specific terms.

Citation

If you use BeyondArena, please cite:

Beyond IID: How General Are Tabular Foundation Models, Really? Lennart Purucker, Andrej Tschalzev, Nick Erickson, Gioia Blayer, David Holzmüller, Alan Arazi, Alexander Pfefferle, Mustafa Tajjar, Gaël Varoquaux, Frank Hutter arXiv:2606.30410

📄 arXiv

BibTeX:

@misc{purucker2026iidgeneraltabularfoundation,
      title={Beyond IID: How General Are Tabular Foundation Models, Really?}, 
      author={Lennart Purucker and Andrej Tschalzev and Nick Erickson and Gioia Blayer and David Holzmüller and Alan Arazi and Alexander Pfefferle and Mustafa Tajjar and Gaël Varoquaux and Frank Hutter},
      year={2026},
      eprint={2606.30410},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2606.30410}, 
}

Per-Dataset References

If you use individual datasets, please also cite their original authors. BibTeX for every dataset in the benchmark is shipped alongside this card in dataset_references.bib (one entry per unique academic_reference_bibtex_key referenced by the dataset metadata files).

Changelog

• [27th May 2026] — Initial release: 142 curated IID and non-IID tasks.