A curated subset of CellxGene (~45M cells) used to align the Stack model after pretraining on full human scBaseCount.
≥ 50,000 cells per dataset
≥ 5 donors per dataset
Author-annotated coarse-grained cell type labels were heuristically identified and transferred to adata.obs["author_cell_type"].
