Cross-Modality TFM Pipeline — Reproducibility Bundle (NeurIPS 2026 submission)

Cross-Modality TFM Pipeline — Reproducibility Bundle

This archive contains the data needed to reproduce the headline results of the NeurIPS 2026 ED Track submission "When Tabular Foundation Models Transfer Across Modalities: A Systematic Evaluation Across 95 Datasets".

Contents

data/ — extracted features (.npz) for the 35 Panel A datasets, one directory per dataset. Each features.npz contains the feature matrix (already passed through the relevant frozen encoder: DINOv2, CLIP, AST, Wav2Vec2, HuBERT, SBERT, ChemBERTa, Phikon, mel-stats, MFCC, or catch22) with train/val/test splits and labels.

results/ — 24 result parquet files containing per-cell scores, baselines, cell verdicts, and aggregate panels. The headline numbers in the paper (94.3% / 77.1% / 96.6% / 91.5%) are computed from panel_a_canonical.parquet and panel_b_canonical.parquet.

Reproducing the headline results

The reproduction scripts and full pipeline code are in the anonymized code repository (URL provided in the supplementary material of the paper). With the code repo cloned and this archive extracted alongside, the headline numbers are reproduced by:

python supplementary/reproduce_main_results.py --quick# Expects:#   Panel A oracle   33/35 = 94.3%#   Panel A deployed 27/35 = 77.1%#   Panel B oracle   57/59 = 96.6%#   Panel B deployed 54/59 = 91.5%

For the full pipeline re-run (ETF training + TabICL inference per dataset), see the README in the code repo.

License

This archive is released under the Creative Commons Attribution 4.0 International License (CC-BY 4.0). Source datasets remain under their original licenses; users are responsible for complying with each upstream dataset's terms when using individual feature files.

Authors

Anonymous Authors (NeurIPS 2026 double-blind submission). Author identities will be disclosed after acceptance.