KAN-Payne: data package for a controlled evaluation of Kolmogorov–Arnold network spectral emulators (version 2)

What is here Every table and figure in the paper is reproducible from these records without retraining. The trained checkpoints are included, so the emulators can also be re-evaluated or redeployed directly. | Directory | Files | Size | Contents ||---|---:|---:|---|| `01_grids/` | 10 | 65 MiB | Frozen per-seed 720/80/200 grids, and the contract assertions verifying that wavelengths, masks, targets and label bounds are byte-identical to the public release. || `02_tuning/` | 28 | 21 MiB | Per-architecture learning-rate × weight-decay tuning, scored on the 80-star validation split only. || `03_benchmark_runs/` | 116 | 518 MiB | The twelve budget-matched formal runs (4 architectures × 3 seeds): training histories, checkpoints, validation and test metrics, and test-set residual arrays. || `04_cost_benchmark/` | 20 | 40 KiB | Training wall-clock, peak GPU memory, inference latency, Jacobian and per-star fit timing (Table 1). || `05_pretrained_audit/` | 6 | 62 MiB | Audit of the public pretrained Payne network on the same split, with its train- and test-set residual arrays. || `06_injection_recovery/` | 185 | 129 MiB | Known-truth injection–recovery: the 200 frozen test spectra injected at S/N = ∞, 100, 50, 20 and refit for all 25 labels. || `07_multistart/` | 66 | 11 MiB | Randomised multistart: the stratified 200-star APOGEE subset and the synthetic 50-star arm. || `08_data_scaling/` | 42 | 226 KiB | Training-set-size axis, *N* ∈ {100, 200, 400, 720}, three seeds, 60 000-update half budget. || `09_loss_ablation/` | 8 | 16 MiB | L1-versus-MSE training-loss ablation for KAN-L. || `10_supplementary_controls/` | 405 | 778 MiB | Large-MLP shape controls, weight-decay sweeps, the new-protocol APOGEE rerun, the three-seed recovery grid, and the activation and spline-grid-range controls, with their independent re-verification. || `11_analysis/` | 257 | 71 MiB | Everything computed *from* the run records: the numbers behind each table, figure and quoted statistic, in four stages. || `12_tables_and_figures/` | 30 | 3 MiB | The six generated LaTeX tables, `PROVENANCE.json` mapping every printed value to its source record, and the manuscript figures. || `13_provenance_and_audit/` | 36 | 12 MiB | Command log, environment capture, code state, source checksums, artifact inventory, and the independent audit: recomputed metrics, numerical-consistency and reproducibility reports, and the test-set leakage audit. || `code/` | 40 | 1 MiB | The analysis, table and figure generators (identical to `analysis/` in the code repository). | Totals and per-file checksums are in `SHA256SUMS.txt` and `MANIFEST.json`.`PATH_MAP.md` maps the published directory names onto the run labels that appear inside the archived records. Verification status All files carry SHA256 checksums (`SHA256SUMS.txt`). Every metric reported in the paper — 142 of 142 — was independently recomputed from the archived arrays and matches. Test-set access controls were independently confirmed: hyperparameter selection and checkpoint selection recorded zero test-set evaluations, and every test evaluation postdates the selection freeze (`13_provenance_and_audit/final_audit/test_leakage_audit.md`).