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

Download TabReD

To download TabReD datasets, follow the steps below.

Step 1. Install uv.

Step 2. Create or log in to your existing Kaggle account and complete the authentication process as described here.

Step 3.

To download all TabReD datasets to the default location, run:

uvx tabred download

To download all TabReD datasets to a custom location, run:

uvx tabred download --output-path path/to/directory

To download only some of the TabReD datasets (for example, the Cooking Time and Weather datasets), run:

uvx tabred download cooking-time weather

Datasets

Dataset	Features	Task	Instances Used	Instances Available	Source
Homesite Insurance	299	Classification	260,753	-	Competition
Ecom Offers	119	Classification	160,057	-	Competition
Homecredit Default	696	Classification	381,664	1,526,659	Competition
Sberbank Housing	392	Regression	28,321	-	Competition
Cooking Time	192	Regression	319,986	12,799,642	Dataset
Delivery ETA	223	Regression	416,451	17,044,043	Dataset
Maps Routing	986	Regression	340,981	13,639,272	Dataset
Weather	103	Regression	423,795	16,951,828	Dataset

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}
}