TabAbax

TabAbax is a pretrained tabular foundation model developed by the Tabular Model R&D Team at BaiRong Intelligent Technology, supporting both tabular classification and regression tasks via a drop-in scikit-learn compatible interface. It delivers competitive in-context prediction performance across public tabular benchmarks, with no task-specific fine-tuning or per-dataset training required.

Implementation Source

All inference code is adapted from the open-source TabICLv2 framework (BSD 3-Clause license). This repository only hosts the released model checkpoints; full inference code, usage examples and provenance documentation are available at:https://github.com/tababax-research/TabAbax

The inference code is derived from TabICL v2.2.0, commit 0dbff3ec8fc68c123c87af77b0ea8b25cd2d23f3. That commit is the version this card compares against. It is not a record of the commit that was copied when the code was first taken from TabICL. See PROVENANCE.md in the code repository for which files match that version and which are modified.

Relationship to TabICL

TabAbax is not an independent implementation from scratch. The TabAbax team modified the architecture and training/prior configuration and trained the released checkpoints.

This repository contains the checkpoints only. The inference code is at https://github.com/tababax-research/TabAbax. That repository does not contain the pre-training code, the synthetic prior, the training datasets, the optimizer configuration, or the checkpoint-generation scripts.

Files

• tababax-classifier.ckpt: classifier checkpoint
• tababax-regressor.ckpt: regressor checkpoint

Intended use

Research and non-commercial evaluation of tabular classification and regression models.

Limitations

Results depend on dataset preprocessing, ensemble settings, hardware, and benchmark version. The reported results should not be interpreted as a guarantee on unseen datasets.

The 2026-10-26 comparison ran on NVIDIA H20. No dataset was skipped. The weights are the renamed files in this repository. Per-split results.pkl files are produced by benchmark/tabular/*/main.py and are not committed; the flattened per-split rows from that run are committed in results/ in the code repository: beyondarena-core2k-2026-10-26-classification.parquet (1446 rows, 98 datasets), beyondarena-core2k-2026-10-26-regression.parquet (554 rows, 44 datasets) and tabarena-v0.1-2026-10-26.parquet (816 rows, 51 datasets). Together the BeyondArena pair is the 2000-split core2k board. The benchmark revisions and software versions from that run were not kept, so a rerun reproduces the procedure rather than these exact scores. See results/README.md in https://github.com/tababax-research/TabAbax.

Citation

If you use this model in your research, please cite our technical report:

@article{cheng2026tababax,
  title={TabAbax: A Tabular Foundation Model Technical Report},
  author={Zhixiong Cheng and Xiaoyun Gao and Shiyuan Han and Jingjing Suo and Zhengli Wang},
  journal={arXiv preprint arXiv:2611.00562},
  year={2026}
}

## Third-party notices

See:
https://github.com/tababax-research/TabAbax/blob/main/NOTICE