Benchmark datasets¶
datasets/ holds the full TabArena-v0.1 suite: 51 curated tabular datasets
(87.4 MB, OpenML suite 457) — 7 regression, 44 classification, 4 small
(<1k rows), 40 medium, 7 large (up to 150k rows), including high-cardinality
categorical sets. Files are parquet with categoricals preserved as strings;
datasets/MANIFEST.md lists per-dataset target, size, license and source.
from mltreelib import list_datasets, load_dataset
list_datasets() # ['APSFailure', 'Bank_Customer_Churn', ...]
ds = load_dataset("churn")
ds["n_rows"], ds["n_cols"] # 5000, 20
ds["dtypes"] # {"state": "categorical", "account_length": "numeric", ...}
ds["na_counts"] # missing shown by default
load_dataset returns the same shape as DatasetBuilder.build(), so it feeds
straight back into the data layer.
BeyondArena
The beyond-IID extension (temporal/grouped splits) ships via the separate
data-foundry package and is intentionally not vendored; see the manifest
for notes.
All datasets in the suite¶
Included verbatim from datasets/MANIFEST.md (single source — the table below
always matches what ships in datasets/):
{% include-markdown "../../datasets/MANIFEST.md" %}