Concept bottleneck models for interpretable variable star classification with Gaia DR3: supplementary datasets, feature matrices, model checkpoints, and source code

This record provides the supplementary datasets, processed concept feature matrices, trained model checkpoints, and full source code underpinning the paper "Concept bottleneck models for interpretable variable star classification with Gaia DR3" (C. Yin, Astronomy & Astrophysics, 2026; article reference aa59990-26). The study introduces the first peer-reviewed application of concept bottleneck models (CBMs) to astronomical source classification. Eight classifiers are benchmarked on about 18,000 Gaia DR3 variable stars across six classes (RRAB, RRC, DCEP, DSCT/GDOR/SXPhe, ECL, MIRA/SR), routing predictions through 12 physically meaningful concepts, and evaluated with five-fold stratified cross-validation, Nadeau-Bengio corrected t-tests and McNemar tests under Holm-Bonferroni control, and an OGLE-IV cross-survey generalization test. Contents: data_products/: five machine-readable tables (CSV) - per-class AUC-ROC; Holm-Bonferroni-corrected significance tests; the accuracy-interpretability Pareto frontier; the 66-pair concept correlation matrix with variance inflation factors; and the cross-survey Kolmogorov-Smirnov statistics. feature_matrices/: the processed 12-concept feature matrices and the five-fold cross-validation split (full 18,000-source sample, CV pool, in-domain test set, and OGLE-IV cross-survey set), the fitted scaler, and the label mapping. model_checkpoints/: trained PyTorch weights (five folds each) for HardCBM, HardCBM-Cal, HardCBM-Linear, SoftCBM, CEM, MLP, and the EndToEndHardCBM prototype. source_code/: the complete Python package, experiment scripts, and configuration files needed to reproduce all tables and figures (fixed global seed 42). The underlying Gaia DR3 and OGLE-IV photometry are publicly available from their respective archives; this record distributes only derived products. See README.md for full details.