lab-medvedeva/GADES2

GADES 2.0 benchmark matrices

Data accompanying the GADES 2.0 paper on GPU-accelerated distance computation for sparse single-cell matrices. The repository holds every matrix used in the timing and accuracy experiments, so that reported numbers can be reproduced without regenerating anything.

Four collections live here.

GeneratedDense

Seventeen dense matrices in CSV format, 70 MB in total. The grid crosses object counts of 10, 100, 1000 and 10000 with feature counts of 10, 100, 1000, 10000 and 100000; combinations that exceed available device memory are omitted. File names follow the pattern <objects>_cells_<features>_features.csv. These matrices drive the dense benchmark and the accuracy comparison against reference implementations.

GeneratedSparse

Seventy five sparse matrices in Matrix Market format, 3.6 GB in total, grouped into fifteen directories by shape. Directory names follow the pattern <objects>_cells_<features>_features. Inside each directory the file name is the density level: 0.5.mtx, 0.75.mtx, 0.9.mtx, 0.95.mtx and 0.99.mtx, where the number is the fraction of zero entries.

Sparsity is the variable that matters most for the algorithm, so this collection is the main source of the scaling curves in the paper.

Real

Twenty two real matrices in Matrix Market format, 3.2 GB in total, split between single-cell assays and recommender system interaction data. The single-cell part covers scRNA-seq and scATAC-seq of human and mouse tissues; the recommender part supplies matrices with a very different shape and density profile, which is what makes the timing comparison honest. Fill rates span three orders of magnitude, from 0.07 per cent in BeerAdvocate to 72.7 per cent in Jester, the only effectively dense matrix in the panel.

Matrix	Domain	Source
PBMC_all, B_T, B_CD8T	scRNA-seq, human immune cells	10x Genomics PBMC reference
PBMC5K, TCells, CellLines, HSC	scATAC-seq	published benchmark collections
Camp, Chen	scRNA-seq, organoid and hypothalamus	original studies
FibrocardRNA	direct reprogramming time course	original study
HLCA_marrow, HLCA_aorta, HLCA_lung	multi-organ atlas	accession GSE109774
MovieLens20M	movie ratings	GroupLens, Harper and Konstan 2015
LastFM	artist listening counts	HetRec 2011
Jester	joke ratings	Goldberg et al. 2001
BeerAdvocate, RateBeer	beer reviews	McAuley and Leskovec 2013
Anime, ModCloth, Pinterest, TaFeng	community ratings, fit feedback, image pins, grocery baskets	see the paper

HugeDatasets

Ten large matrices in AnnData format (.h5ad), five single-cell atlases and five from other domains, stored so that they can be read in batches without loading the whole matrix into memory. TabulaSapiensV1 is the whole human Tabula Sapiens atlas, release 1.0, TabulaMuris holds the 3-month-old mice of Tabula Muris Senis, MouseAtlas is the mouse sci-ATAC-seq atlas of chromatin accessibility, and FibrocardATAC is single-nucleus chromatin accessibility of the adult human heart (GEO accession GSE165837).

File	Cells	Features	Non-zero entries
TabulaSapiensV1.h5ad	483,152	58,870	1,271,192,991
TabulaMuris.h5ad	90,120	17,983	214,912,538
MouseAtlas.h5ad	81,173	436,206	421,971,103
FibrocardATAC.h5ad	79,515	287,415	138,840,495
AIDA.h5ad	1,058,909	35,240	2,145,186,863
NYTimes.h5ad	300,000 documents	102,660 words	69,679,427
Netflix.h5ad	480,189 users	17,770 movies	100,480,507
Pokec.h5ad	1,632,803 users	1,632,803 users	30,622,564
BookCrossing.h5ad	278,858 users	271,379 books	1,031,175
AmazonVideoGames.h5ad	2,766,656 users	137,249 products	4,555,500

AIDA holds the raw counts of the Asian Immune Diversity Atlas release published on CZ CELLxGENE Discover (dataset c51e913d-832a-44a1-8f43-d20c51a1799f): the matrix is taken from raw/X of the source file, cell metadata is dropped, and cell barcodes and gene identifiers are kept as names. The source is distributed under CC BY 4.0, so reuse requires citing the AIDA consortium. Its 2,145,186,863 non-zero entries sit just below the 32-bit index limit of 2,147,483,647.

TabulaSapiensV1 holds the raw counts of Tabula Sapiens release 1.0 (The Tabula Sapiens Consortium, Science 2022, doi 10.1126/science.abl4896), taken from the file TabulaSapiens.h5ad of figshare item doi 10.6084/m9.figshare.14267219.v5. The source file carries four matrices of the same shape: normalized values in X, raw counts in raw/X and in the raw_counts layer, and ambient-RNA-corrected counts in the decontXcounts layer. Only the raw_counts layer is kept, because the GADES analyses were computed on raw counts. Cells from both 10x (456,101) and Smart-seq2 (27,051) are included. TabulaMuris holds the raw counts (raw/X) of the 3-month-old mice in Tabula Muris Senis as published on CZ CELLxGENE Discover (dataset 4e0a8071-1b1d-49d0-a5ac-6f4f5c01e99f; Tabula Muris Consortium, Nature 2020, doi 10.1038/s41586-020-2496-1): 90,120 cells, of which 45,602 come from 10x and 44,518 from Smart-seq2. Both sources are distributed under CC BY 4.0, so reuse requires citing the respective consortium.

NYTimes holds word counts of New York Times articles from the UCI Bag of Words collection, with documents named doc1 onward in the original order and words taken from the collection's vocabulary. Netflix holds the ratings of the Netflix Prize training set: users are renumbered consecutively and keep their original identifiers as names (user6 and so on), movies keep their original numbers (movie1 to movie17770), and values are ratings from 1 to 5. BookCrossing holds ratings from the Book-Crossing community (Ziegler et al., 2005). The shape follows the user and book catalogues of the original release, so 118,605 ratings on ISBNs missing from the book catalogue are left out. In the original data a rating of 0 marks an implicit interaction, which a sparse matrix cannot tell apart from an empty cell, so every rating is shifted by one: implicit interactions become 1 and explicit ratings 1 to 10 become 2 to 11. Users are named user plus their original identifier and books by ISBN. AmazonVideoGames holds ratings from the Video Games category of Amazon Reviews 2023 (Hou et al., ACL 2026), built from raw/review_categories/Video_Games.jsonl at Hugging Face revision 2b6d039ed471f2ba5fd2acb718bf33b0a7e5598e (SHA-256 7bde9e0a039bdb321e993bfe78a6c7d438fb1504b1ad152c0cb6a33a8b7cb26a). Products are identified by parent_asin, values are ratings from 1 to 5, and when a user reviewed the same product several times the latest review is kept. Users and products keep their original identifiers as names. Pokec is the friendship graph of the Slovak social network Pokec from the SNAP collection: the graph is directed, row i lists the users that user i named as friends, and entries are 1.

Each file follows the AnnData convention: X is a CSR matrix of cells by features, so the first dimension is always cells. The matrices used elsewhere in this repository are stored features by cells, which is why these atlases were transposed during conversion. obs holds a n_nonzero column with the number of non-zero entries per cell, and var holds a n_cells column with the number of cells in which each feature is non-zero. FibrocardATAC carries the original cell barcodes and peak coordinates written as chr:start-end; TabulaSapiensV1, TabulaMuris and AIDA keep the cell and gene identifiers of their sources; the source matrix of MouseAtlas carries no names, so its cells and features are named by position (cell0, feature0 and so on).

The files open with anndata.read_h5ad, including backed="r" mode, and with the streaming reader shipped with GADES, which reads a slice of cells at a time through the cached indptr. Every file was checked against its source matrix cell by cell on a random sample before upload.

Provenance, licensing and takedown

Every matrix here is a processed form of a published dataset, and the original terms of each source apply to it. Full citations are given in the paper and in its supplementary table; this card names sources so that each matrix can be traced back.

Four points deserve to be stated plainly rather than buried.

• BeerAdvocate and RateBeer are no longer distributed by their original holders. They are included here because the paper reports timings on them and those numbers should stay checkable, but anyone reusing them should consult the original authors first.
• Netflix Prize data was withdrawn from public distribution by Netflix. The matrix is kept here because the paper reports timings on it, and it should not be reused as a substitute for the original release.
• Amazon Reviews 2023 is published without a stated license; its authors describe it as made available primarily for research purposes. The Video Games matrix is included here for research reproducibility only.
• MovieLens-20M is distributed by GroupLens under terms that restrict redistribution. The copy here is a processed interaction matrix retained for reproducibility of the reported timings, not a substitute for obtaining the dataset from GroupLens.

If a rights holder objects to any matrix in this repository, write to the lab and it will be removed.

Citation

If these matrices are used, cite the GADES 2.0 paper and the original source of every matrix that is touched. Citing this repository alone is not enough: the underlying data belongs to the studies that produced it.