Class-imbalance remedies for tabular foundation models: historical and supplementary probabilities and experiment records

Abstract

Version 2.0.0 provides two separate collections of 8,600 persisted probability files each, covering historical and supplementary experiments on the same 43 binary tasks and five seeds. Each NPZ contains validation/test probabilities and aligned binary labels (p_val, p_test, y_val, y_test). No raw feature tables or model weights are included. The four original proba_part ZIPs are inherited unchanged and contain 8,600 historical files under results/proba/. The five additional supplement_part01.zip through supplement_part05.zip archives together contain 8,600 new probability files and 8,600 cell JSON records under revision/supplement/, 860 forty-trial parameter-search records, 215 split-index files, and the historical 60,200-row seven-workflow score table. All five supplementary ZIPs are required for the complete collection. Identically named historical and supplementary probabilities must not be merged into one directory. Cache paths in new cell metadata are normalized; both original and released JSON hashes are retained. The companion software version provides all analysis code, contrast tables, model/sampler provenance, and corrections to v1.x interpretations. In particular, this is a workflow-dependent empirical evaluation, not evidence of a universal failure of remedies for all TFMs. New TabDPT comparisons use seeded same-run baselines; the old fit-time PCA RNG cannot be recovered. Threshold/calibration, source-dependence and protocol limitations are documented in the code erratum. Original generated artifacts are CC BY 4.0. Upstream datasets and weights retain their separate terms. README-DATA.md explains the file layout; SHA256SUMS.txt records archive checksums. The code-side installer verifies archive and individual-file hashes before extraction.