actinn-jax pre-trained cell-type annotation references (human and mouse)

Pre-trained cell-type annotation references for actinn-jax, a fast CPU-only reference-mapping classifier for single-cell RNA-seq. Each archive unpacks to a HierarchicalReferenceModel directory loadable with actinn_jax.bundled_reference(name), which downloads and caches them automatically on first use. broad_human_v1 — census-wide human, ~800 cell types / 314 tissues; coarse hierarchy from scPRINT embeddings. broad_mouse_v1 — census-wide mouse, 453 cell types / 85 tissues; coarse hierarchy from Cell Ontology lineage (no GPU in the build). On two datasets held out of the reference entirely: 0.638 ontology-aware concordance, 0.718 over the 71% of cells kept at min_prob=0.5. panhuman_distill_v1 — human, 324 cell types, distilled from Pan-human Azimuth; on a withheld liver study it scores 0.406 ontology concordance against 0.338 for broad_human_v1, at roughly three times its throughput. liver_hlica_v1 / v2 — focused human liver references (38 and 48 types) from the Human Liver Cell Atlas, for the refinement tier of the workflow. All references are built from CELLxGENE Census (release 2025-11-08). Build scripts and the benchmark behind them: actinn-jax-benchmark. Attribution. panhuman_distill_v1 is derived from Pan-human Azimuth (Sarkar, Li, Molla, ... Satija, bioRxiv 2026, doi:10.64898/2026.07.16.738997), whose model weights are © the authors under CC BY 4.0 (doi:10.5281/zenodo.20401417), obtained via the MIT-licensed panhumanpy package. The liver references derive from the Human Liver Cell Atlas (Edgar et al. 2026, CC BY 4.0, doi:10.64898/2026.06.30.735539).