SCALAR: cross-scale nanoparticle benchmark generation pipeline and evaluation code

Code and generation pipeline for the SCALAR cross-scale nanoparticle benchmark. Phase I builds nanoparticles from unit-cell CIFs: primitive-cell setup, supercell replication to at least 2Rmax + δbox in every direction, carving-centre selection, spherical truncation for R = 10–30 Å, and box-margin verification. Phase II samples rigid rotations on SO(3) under a geodesic-spacing constraint. Phase III applies split-aware partitioning by carving radius. Also included: the language-model evaluation tasks (direct property prediction, chain-of-thought variants, inverse retrieval), graph-neural-network baselines, and the dataset characterisation analyses. Regenerating all 83 materials at all 21 radii reproduces the archived seed structures' atom counts exactly for 1,737 of 1,743 structures, with coordinates agreeing to 10-8 Å or better for 82 of the 83 materials. The seed inputs are archived separately; the concept DOI 10.5281/zenodo.20631920 resolves to the latest version of that record.