## A Living Benchmark for Machine Learning on Tabular Data π«
---
| π [Leaderboard](https://tabarena.ai/) | π [Example Scripts]( https://tabarena.ai/code-examples) | π [Dataset Curation](https://tabarena.github.io/data-foundry/) | π Papers: [TabArena-v0.1](https://arxiv.org/abs/2506.16791) Β· [BeyondArena](https://arxiv.org/abs/2606.30410) |
|:--------------------------------------:|:----------------------------------------------------------------------------------------:|:----------------------------------------------------------------------------------------:|:--------------------------------------------------------------------------------------------------------------:|
---
TabArena is a living benchmarking system that makes benchmarking tabular machine learning models a reliable experience. TabArena implements best practices to ensure methods are represented at their peak potential, including cross-validated ensembles, strong hyperparameter search spaces contributed by the method authors, early stopping, model refitting, parallel bagging, memory usage estimation, and more. Explore the latest results on the [live leaderboard](https://huggingface.co/spaces/TabArena/leaderboard).
This single codebase powers **two complementary benchmarks** that share the same fitting, runner, and
evaluation code:
- ποΈ **TabArena-v0.1** β the living benchmark on **curated, IID** tabular datasets.
- π **BeyondArena** β a holistic, ***beyond-IID*** benchmark spanning **IID, temporal, and grouped**
tasks across a wide range of dataset sizes and feature dimensionalities.
**BeyondArena will superseed TabArena-v0.1 in the future.**
> [!TIP]
> **New here? Start with TabArena, then graduate to BeyondArena.** Get your model working and
> competitive on TabArena's curated IID datasets first; once it holds up there, run the *same* code
> on BeyondArena to stress-test how well it generalizes beyond IID.
**TabArena** covers 51 curated datasets (9β30 splits each) and 27+ methods, including 10+ tabular
foundation models β over 50M trained models, with all validation and test predictions cached for
tuning and post-hoc ensembling. **BeyondArena** extends this to **[142 datasets](https://huggingface.co/datasets/TabArena/BeyondArena)** across IID,
temporal, and grouped task types, spanning tiny to 1M-row datasets and low- to high-dimensional
features.
## β‘ Quickstart
> [!TIP]
> The fastest way to try TabArena end-to-end:
```bash
pip install uv
git clone https://github.com/autogluon/tabarena.git && cd tabarena
uv venv --seed --python 3.12 && source .venv/bin/activate
uv pip install --prerelease=allow -e "./packages/tabarena[benchmark]"
python examples/benchmarking/run_quickstart_tabarena_model.py # benchmark a model that TabArena tunes
python examples/benchmarking/run_quickstart_tabarena_system.py # benchmark a system that tunes itself
```
TabArena ranks **models** (one method, tuned by TabArena under a shared protocol) and **systems**
(a pipeline that does its own tuning and ensembling, like AutoGluon); see
[Contributing a Model or System](#contributing-a-model-or-system) for the difference and how to submit yours.
For other install paths (eval-only, editable AutoGluon, dependency), see [Installation](#-installation) below.
To try **BeyondArena** instead, run `python examples/beyondarena/run_quickstart_beyondarena_model.py`
(or `run_quickstart_beyondarena_system.py`) with the same install.
## πΉοΈ Use Cases
We share more details on various use cases of TabArena in our [examples](examples):
* π **Benchmarking Beyond IID (BeyondArena)**: please refer to [examples/beyondarena](examples/beyondarena).
* π **Benchmarking Predictive Machine Learning Models and Systems**: please refer to [examples/benchmarking](examples/benchmarking); to get yours onto the leaderboard, see [Contributing a Model or System](#contributing-a-model-or-system).
* π§ͺ **Advanced and Specialized Usage** (incl. using a TabArena model directly on your own data): please refer to [examples/advanced](examples/advanced).
* ποΈ **Analysing Metadata and Meta-Learning**: please refer to [examples/meta](examples/meta).
* π **Generating Plots and Leaderboards**: please refer to [examples/plots](examples/plots).
* π’ **Per-dataset and per-split results**: not hosted. Every leaderboard script in [examples/plots](examples/plots) writes them as `results_per_split.csv` (one row per dataset, split and method) next to its figures; [`run_export_results_per_split.py`](examples/plots/run_export_results_per_split.py) writes the CSVs without the figures.
* π **Reproducibility**: we share instructions for reproducibility in [examples](examples).
### Datasets
Please refer to [Data Foundry](https://github.com/TabArena/data-foundry) ([documentation](https://tabarena.github.io/data-foundry/)) to learn more about the datasets or to contribute data.
### Contributing a Model or System
TabArena accepts two kinds of entrant: a **model** (one method that TabArena tunes under its shared
protocol) and a **system** (a pipeline that owns its own preprocessing, validation, tuning and
ensembling inside the budget TabArena hands it). TabArena is not a benchmarking service: evaluate your
method on TabArena-Lite first, open a pull request with the template, and a maintainer verifies and
re-runs it for the final entry. The details:
