Paper and reproducibility

SoftwareX paper

TabuLLM is described in:

Sharabiani, M.T.A., Bottle, A., and Mahani, A.S. (2026). TabuLLM: Feature extraction and interpretation of text columns in tabular data using large language models. SoftwareX, 35, 102810. https://doi.org/10.1016/j.softx.2026.102810

Reproducibility assets

The paper_experiments/ folder in the repository contains the full reproduction package for the fraud-detection benchmark experiments reported in the paper:

  • reproduction instructions, covering everything from regenerating figures from committed summaries to a full from-scratch rerun of the experiment matrix

  • paper-requirements.txt, a pinned pip freeze of the exact environment used for the manuscript runs

  • run scripts for the two benchmark sweeps (k_sweep/, col_sweep/) and the summarisation/plotting scripts that turn their output into the manuscript figures

  • archived results (per-fold predictions, metrics, and rolled-up summaries) committed to the repository, plus pre-computed embeddings and result bundles archived on Zenodo (see the reproduction instructions for record details)

The documentation build does not execute the paper experiments; see the reproduction instructions above to rerun them.

Citation

If you use TabuLLM in academic work, please cite:

Sharabiani, M.T.A., Bottle, A., and Mahani, A.S. (2026). TabuLLM: Feature extraction and interpretation of text columns in tabular data using large language models. SoftwareX, 35, 102810. https://doi.org/10.1016/j.softx.2026.102810