TabCausal

TabCausal is a tabular causal discovery model for predicting directed causal graphs from tabular data.

This repository hosts the released TabCausal model checkpoint. The source code, inference scripts, benchmark utilities, and examples are available at:

https://github.com/LAMDA-Tabular/TabCausal

Checkpoint

checkpoints/tabcausal-base.pt

Installation

git clone https://github.com/LAMDA-Tabular/TabCausal.git
cd TabCausal
pip install -r requirements.txt

Usage

python -m tabcausal.cli predict \
  --checkpoint checkpoints/tabcausal-base.pt \
  --input /path/to/data.npz \
  --output outputs/prediction.npz \
  --device cuda:0

The output file contains directed-edge logits, probabilities, and a predicted adjacency matrix.

TabCausal supports benchmark-style .npz files and common numeric tabular formats such as .csv, .tsv, .npy, .parquet, and .pkl.

Benchmark Evaluation

Benchmark generation and evaluation scripts are included in the GitHub repository. Please refer to the GitHub README for detailed instructions.

Citation

If you use TabCausal in your research, please cite:

@article{li2026tabcausal,
  title={TabCausal: Pretraining Across Causal Environments for Tabular Causal Discovery},
  author={Li, Zi-Rong and Liu, Si-Yang and Wang, Tian-Zuo and Ye, Han-Jia},
  journal={arXiv preprint arXiv:2605.31156},
  year={2026},
  url={https://arxiv.org/abs/2605.31156}
}