tabred 0.1.2

TabReD: Analyzing Pitfalls and Filling the Gaps in Tabular Deep Learning Benchmarks

TabReD is a collection of eight industry-grade tabular datasets designed to evaluate machine learning methods under more realistic conditions: temporal distribution shift, rich feature sets from feature engineering pipelines, closer aligned with some real world applications of tabular machine learning.

📜 arXiv 📚 Other tabular DL projects

Datasets

Preprocessed data format

The downloader unpacks each dataset into its own directory:

data/<dataset>/
├── info.json
├── x_num.npy
├── x_cat.npy          # when present
├── x_bin.npy          # when present
├── x_meta.npy
├── y.npy
└── splits/
    ├── default/
    │   ├── train.npy
    │   ├── val.npy
    │   └── test.npy
    ├── random-{0,1,2}/
    │   ├── train.npy
    │   ├── val.npy
    │   └── test.npy
    └── sliding-window-{0,1,2}/
        ├── train.npy
        ├── val.npy
        └── test.npy

The x_*.npy files contain feature matrices, y.npy contains targets, and info.json contains task metadata. Split files contain row indices into these arrays. The default split is the main split from the TabReD paper; the random and sliding-window splits are provided for split-strategy studies.

Repository structure

• src/tabred: downloader package and CLI.
• paper: code for reproducing the paper

Rebuilding datasets from raw sources

Most users should use the preprocessed downloader above. The preprocessing pipeline is kept in this repository for reproducibility and maintenance, but it is not the recommended way to obtain the benchmark data.

The cleaned-up dataset preparation scripts will be published in a few days (@puhsu 10.07.26). For now consult the previous commits and the preprocessing folder there.

Citation

If you use TabReD, please cite:

@inproceedings{
 rubachev2025tabred,
 title={TabReD: Analyzing Pitfalls and Filling the Gaps in Tabular Deep Learning Benchmarks},
 author={Ivan Rubachev and Nikolay Kartashev and Yury Gorishniy and Artem Babenko},
 booktitle={The Thirteenth International Conference on Learning Representations},
 year={2025},
 url={https://openreview.net/forum?id=L14sqcrUC3}
}