junidude14/PBISC-data

PBISC Diffusion training data

Private archival/reproduction release for PBISC Diffusion. The catalog preview lists shards; it is not the expression matrix or a train/test example split. Expression remains unchanged sparse CSR Zarr (base v3, overlay v2) in transport archives. Use the PBISC code package to verify and restore assets.

The release has 41,395,184 logical cells, 22 source datasets, 432 shards and 20,097 ordered genes. It includes row-aligned cell metadata, measurement masks, donor splits, sealed E2 recipes, All-K top-ups and the train-fitted conditioner. Canonical counts are shared across model panels. Measured zero and unmeasured coordinates remain distinct. Source assay units and sparse values are preserved.

release.json is the complete asset/hash manifest. relocation.json binds new paths while retaining original source/index fingerprints. Original manifests are preserved rather than rewritten. No new random split is created.

Public redistribution rights have not been established by this export. provenance/rights.tsv records each study as unverified. Private archival storage does not establish a blanket dataset license. No license is inferred from public accessibility of an upstream download. This dataset remains private pending a source-specific rights review.

Code: https://github.com/junidude/PBISC Model: https://huggingface.co/junidude14/PBISC

Verify and restore pinned assets with pbisc data download or pbisc data restore. These archives are not directly loaded as dense matrices by load_dataset(). Plan approximately 112 GB for transport plus a similar amount for full extraction; selected-shard restoration is supported. The current release is file/provenance packaging, not a new biological validation of the source datasets.