Specimen-level generalization of handheld LIBS machine learning classifiers for lithium ore characterization: data, code, and source data

Data, analysis code, and per-figure source data supporting the article "Specimen-level generalization of handheld LIBS machine learning classifiers for lithium ore characterization" (submitted to Spectrochimica Acta Part B: Atomic Spectroscopy). CONTENTS 810 handheld laser-induced breakdown spectroscopy (LIBS) spectra acquired from 52 museum specimens of five lithium-bearing pegmatite phases (spodumene, lepidolite, amblygonite, petalite, triphylite) using a SciAps Z-300 analyzer under argon purge, together with the complete preprocessing and analysis pipeline used in the article. The archive supports the article's central methodological result: that spectrum-level cross-validation, in which replicate spectra from a single physical specimen are split across training and test folds, overestimates deployment-relevant classification performance relative to specimen-level validation. FILES - README.md — description, setup instructions, and contents- libs-specimen-level-v1.0.zip (3.5 MB) — code, per-figure source data, figures, documentation, and metadata arrays- spectra_raw.npz (67 MB) — raw instrument intensities (810 x 23,431 channels, 180-961 nm), float32, row-aligned with manifest.csv REPRODUCIBILITY Two utilities allow the reported results to be checked independently: - verify_canonical.py compares every value reported in the article against the deposited source data (36 checks).- replot_all_figures.py regenerates every figure from the deposited source data, so figures can be restyled or re-exported without repeating the analysis. Both run from the 3.5 MB archive alone. The full analyses (generalization-gap grid, provenance classification, and the leak-free Leave-One-Specimen-Out evaluation) additionally require spectra_raw.npz. See docs/02_CAPABILITY_MATRIX.md for which command reproduces which result. Environment: Python 3.12, scikit-learn 1.4, numpy, scipy, pandas, matplotlib. Global random seed 42. LICENSING Code is released under the MIT License; data, source data, and figures under CC BY 4.0. See LICENSE-CODE.txt and LICENSE-DATA.txt in the archive. Raw per-shot spectral files are available from the corresponding author on reasonable request.