theislab/cellxgene-scbasecount-human-tokenized

CellxGene + scBaseCount, human, tokenized

Geneformer-style rank-value tokenized human cells: a straightforward union of two existing public corpora, CellxGene Discover (cxg_raw_2025-11-08) and scBaseCount (2025-02-25), deduplicated (CellxGene's own is_primary_data flag, to drop repeated cells across dataset views) and tokenized. 210,202,463 cells from 2,225 source files. This is not an original data-collection effort -- all underlying cells were already public via CellxGene Discover and scBaseCount; this repo is just that union in a ready-to-train tokenized form.

Updates

2026-11-03 update (v2 rebuild): the corpus was rebuilt to fix three problems found in the previous version:

• Medians. The previous gene-medians file had 7 genes at exactly 0 (forcing a division by zero) and 3,762 more set to a single filler value, so those genes ranked arbitrarily. Medians are now recomputed from this corpus's own raw counts (Geneformer recipe: per cell, non-zero counts normalized to a sum of 10,000; per gene, the median of its non-zero values over every cell), over all ~2.1x10^11 values, using a histogram-based median to avoid holding them in memory.
• Wrong matrix in some source files. Some h5ads store log-normalized values in X and raw counts in raw/X. An audit pass now picks the raw-count matrix per file (checked to be non-negative and integer on every stored value) instead of assuming X; files that fail the check are excluded rather than silently tokenized from the wrong matrix.
• No counts. Every token now keeps its raw UMI count (new counts column), so count-level augmentations (read-depth thinning, ambient-RNA injection) are possible at train time. A pooled ambient ("soup") profile per source file is included for this.

Rows are globally shuffled, as of the 2026-10-19 update below. Checked by a whole-corpus QC pass (0 violations of every per-cell structural rule over all 210,202,463 cells) and an independent rebuild of 1,000 random cells straight from the source h5ads with plain anndata, which matched the pipeline's own output 1000/1000. The build scripts, with a full write-up of the fix, the checks, and their results, are at scripts/gene_medians/ in the theislab/leCellJepa repo.

Known gap: 3,216 of 23,262 vocab genes never occur in the corpus (no cell in it expresses them above zero), so they have no median and never appear as a token. An eval tokenizer using this vocab on a dataset that contains one of them should skip it, the same way training does, rather than raising a key error.

2026-10-19 update: rows are now globally shuffled (previously contiguous by source study -- a streaming data loader without its own shuffle would see batches, and a leading-rows validation split, dominated by whichever study happened to be first), and 6 genes whose corpus median was exactly 0 have been dropped from every cell's token list (count / median = inf was forcing them to a spurious rank-1 position in ~1.8% of cells). Cell count and the rest of each cell's ranking are unaffected. (Superseded by the 2026-11-03 update above, which recomputes every gene's median rather than just dropping the 6 worst.)

Contents

• Repo root -- a HuggingFace datasets DatasetDict (single train split) saved via save_to_disk (dataset_dict.json, train/). Columns: input_ids (large_list, Geneformer-style rank-value encoding, <cls>-prefixed and <eos>-terminated, truncated to 4096 tokens), counts (large_list, raw UMI count per token, 0 at the special tokens, clipped at 65,535), n_counts (int32, the cell's total raw count over all genes), n_genes (int32, non-zero genes with a median, before the 4094-gene cut), cell_index (row in the source h5ad), source_id (the source file's id).
• vocab/hsapiens_token_dictionary.pkl -- the exact vocab these ids were tokenized against: 23,266 entries, <pad>=0, <mask>=1, <cls>=2, <eos>=3, genes from id 4 (23,262 genes). This is not the same as the default packaged token_dictionary.pkl (~20k genes) -- loading with the wrong vocab silently misaligns gene ids. Changed from the previous version: that one had already pruned the 6 zero-median genes out of the vocab itself (23,260 entries); v2 does not prune the vocab, since a gene with no median is simply never emitted as a token (see below).
• vocab/hsapiens_gene_medians.pkl -- non-zero median gene expression values used by the tokenizer's normalization step. Replaced in the 2026-11-03 update: recomputed from this corpus's own raw counts rather than an earlier, under-sampled file.
• soup/soup_matrix.npy + soup/soup_index.json -- one row per source file (float32, rows sum to 1): that file's pooled raw-count profile over its kept cells, for ambient-RNA-injection augmentation. soup_index.json maps source_id to row index.
• qc/report.json -- the whole-corpus QC summary from the 2026-11-03 rebuild.

Loading

from huggingface_hub import snapshot_download
from datasets import load_from_disk

local_dir = snapshot_download(repo_id="theislab/cellxgene-scbasecount-human-tokenized", repo_type="dataset")
ds = load_from_disk(local_dir)["train"]