PubMed 2026 MeSH-Incidence Corpus: a sparse-binary dataset (30 M articles × 30 k descriptors) for self-organizing-map and clustering benchmarks

A MeSH-descriptor incidence matrix derived from the PubMed 2026 annual baseline. Each row is one indexed article and each column is one Medical Subject Headings (MeSH) descriptor; an entry marks that the article was indexed with that descriptor. The result is a large, genuinely sparse binary matrix (~11 descriptors per article) intended for benchmarking self-organizing maps, clustering, and sparse-matrix kernels at scale. Scale. 29,903,261 articles × 30,766 MeSH descriptors, 332,436,043 non-zeros (mean 11.1 descriptors/article), after dropping articles with fewer than 5 descriptors. Contents. corpus.sbcsr — the full incidence matrix, unsplit (CSR binary; see format below) vocab.json — the column→MeSH map: an ordered list of MeSH descriptor UI codes (e.g. "D000123") summary.json — provenance: counts, the ≥5-descriptor filter, source filenames, timestamps README.md — this record, the format spec, and regeneration instructions Splitting. The corpus is shipped unsplit so split ratios and seeds can be changed in code without re-publishing the data. To reproduce the SparseSOM Phase 1 experiments, the partition is train/held-out = 90/10 at seed 42, generated by the reproduction repository's scripts/split_corpus.py (--seed 42 --train-frac 0.9); this deterministically regenerates the paper's exact partition (train 26,912,934 articles, held-out 2,990,327), and the repository's pipeline performs and verifies this split automatically. What it deliberately does not contain. No abstracts, no titles, no author or journal data, and no PMIDs — rows carry no identifiers of any kind. Rows appear in natural document order, which is reproducible from the public MEDLINE baseline, so the row→article correspondence can be re-derived by replaying the pipeline over the public source data; this is stated for transparency and has no confidentiality implication, as both PMIDs and MEDLINE are public. Only NLM's controlled-vocabulary descriptor codes are stored (not even the descriptor strings). The bundle is therefore composed solely of public-domain, U.S.-government-produced content. Format (.sbcsr, little-endian). A 24-byte header — magic "SBCSR1\0\0", then uint32 n_samples, n_features, n_nonzeros, reserved — followed by uint32 row_ptr[n_samples+1] and uint16 col_idx[n_nonzeros]. Column id c maps to MeSH UI vocab.json[c]. License. Released CC0 1.0 (public-domain dedication) on these derived structures. Attribution (required). Courtesy of the U.S. National Library of Medicine. This product uses publicly available data from NLM but is not endorsed or certified by NLM. Currency. This is a static snapshot of the PubMed 2026 annual baseline and does not reflect the most current or accurate data available from NLM. For live data, ingest directly from https://ftp.ncbi.nlm.nih.gov/pubmed/baseline/. Integrity. medline-mesh-pubmed26-ge5.tar.gz — SHA-256 e0070e090a884ef1069c0c97b9baba300d89373938c658a3d7b6b8322ed4f919. Changes in this version. Data files are bit-identical to v1 (corpus.sbcsr, vocab.json, summary.json unchanged). The bundled README and this description are corrected: the record previously misdescribed the bundle as pre-split (it is shipped unsplit; the paper's split is 90/10 at seed 42 as above) and overstated row anonymity as untraceable.