Dataset Announcements

Coyo-VE is a multimodal embedding dataset in which each vector is a joint image-text embedding generated using Qwen3-VL-Embedding-2B. The source data is drawn from the COYO subset of the larger LLaVA-OneVision-1.5-Mid-Training-85M dataset.

The benchmark includes 25,000 queries with exact ground-truth results at k = 1000.

License: CC BY 4.0.

PubMed-VE is an embedding dataset generated from the collection of PubMed abstracts available through the MedRAG-PubMed dataset using BGE-M3. It includes dense, sparse, and multi-vector representations.

The benchmark includes 1,000 queries per vector type with exact ground-truth results at k = 1000.

License: Apache License 2.0.

DINO is a public dense-vector benchmark containing image-patch embeddings extracted from YFCC100M with the DINOv3 ViT-L/16 model. The embeddings are quantized to uint8 and use Euclidean distance. The framework includes 13 registered scales from 100,000 to 2 billion base vectors.

License: The embedding vectors and auxiliary files are released under CC BY-NC 4.0. The original YFCC100M images retain their respective Creative Commons licenses.

Pes2o-VE is an embedding dataset generated from more than eight million academic papers in the Pes2o text corpus. It uses a combination of recursive splitting and semantic chunking with the Qwen3-Embedding-4B model.

The benchmark includes approximately 20,000 queries with exact ground-truth results at k = 10.

License: ODC-BY 1.0.

Q-FineWeb-10B is an embedding dataset generated from the FineWeb text corpus using GTE-Multilingual-base. It provides both dense and sparse embeddings at a scale of more than ten billion vectors.

The benchmark includes 100,000 dense queries, 10,000 sparse queries, and approximately 10,000 filtered queries, each with exact ground-truth results at k = 1000.

License: ODC-BY 1.0.

arxiv-for-fanns is a filtered approximate nearest neighbor search benchmark containing 4096-dimensional embeddings of arXiv paper abstracts generated with the Stella 400M embedding model. Each vector is accompanied by 11 attributes, including update date, authors, and category, derived from the Cornell University arXiv dataset.

• Small: 1,000 items, 197 MB.
• Medium: 100,000 items, 1.86 GB.
• Large: More than 2.7 million items, 46.1 GB.

License: Not specified in the submission.

MoReVec is a relational filtered approximate nearest neighbor benchmark with two joinable tables: 551,155 movies and 2,598,267 reviews at its largest scale. It contains L2-normalized, 768-dimensional text embeddings generated from real movie synopses and reviews using gte-base-en-v1.5.

Each scale includes numeric and categorical metadata, selectivity filter banks, and exact 100-nearest-neighbor ground truth. The large corpus also provides hardness-controlled hard and superhard query packs with 1,000 queries per table, including Match-PDF workloads on α-Hardness and Post_Hardness and a near-duplicate density exploit, generated using HCBGen.

License: CC BY 4.0.

PLASMA provides pre-built graph indices that enable users to evaluate memory-layout effects across diverse graph structures on their own GPUs. It includes indices built by CAGRA, NSG, NSSG, NN-Descent, and DiskANN/Vamana over 14 dataset configurations at scales from 1M to 50M, with permutation maps for seven vertex orderings: GOrder, RCM, Hub Sort, Hub Cluster, in-degree, out-degree, and random.

The configurations span classical descriptors (SIFT, GIST, and Deep) and modern embeddings (Yandex T2I, OpenAI, Wikipedia, BioASQ, and C4), ranging from 96 to 1536 dimensions. Providing graph structures and orderings separately allows topology and memory layout to be varied without rebuilding the indices.

License: The graph structures, authored by Yutaro Oguri, Mai Nishimura, and Yusuke Matsui, are released under CC BY 4.0. This license does not cover the underlying vectors or source corpora. Users must obtain vectors from their original providers and comply with the applicable terms, which are itemized on the dataset page.

These benchmarks contain embeddings of raw genomic sequence reads from the NIH Sequence Read Archive and test vector search databases on large-scale sequence search. The embeddings were generated using LOCALE, a locale-alignment embedding model designed for noise-robust DNA search.

• locale-benchmark-sra50: a 50-accession, 30 GB benchmark.
• locale-benchmark-sra500: a 500-accession, 500 GB benchmark.

License: CC BY 4.0.

Caselaw is a public vector benchmark based on the Collaborative Open Legal Data Cases collection derived from CourtListener bulk data. It contains legal opinions and rich metadata such as filing date, court type, court name, and jurisdiction. Opinions are embedded with OpenAI's text-embedding-3-small model; long opinions are represented as multiple vectors to accommodate documents exceeding the model's context window.

The release includes a multi-vector binary format, metadata and curated filter queries, and exact top-100 nearest-neighbor ground truth computed with the Chamfer aggregation metric and Euclidean distance. Filtered ground truth is provided for the full base set and smaller prefixes. A separate single-vector release includes documents that fit within one embedding, along with workloads grouped into low, medium, and high filter match-rate regimes.

License: The embeddings and auxiliary files are released under CDLA Permissive 2.0. The original COLD Cases dataset is released under CC0 1.0 Universal.

YFCC-Images is a public vector benchmark derived from the image subset of YFCC100M. Each image is represented by a normalized, 1280-dimensional embedding generated with the CLIP-based CLIP-ViT-bigG-14-laion2B-39B-b160k model. It expands upon the YFCC benchmark released for the NeurIPS 2023 Practical Vector Search Challenge, but does not include the additional metadata required for filtered search.

The release uses the standard single-vector fbin format and includes exact top-100 nearest-neighbor ground truth for the full base set and both subsets. Ground truth was computed using Euclidean distance and, because the embeddings are unit-normalized, is also applicable to cosine similarity and inner-product search.

License: The embeddings and auxiliary files are released under CDLA Permissive 2.0. The original images retain their respective Creative Commons licenses, and the original metadata is subject to the Yahoo Webscope License. This release does not redistribute the source images or metadata.

In-distribution datasets

Out-of-distribution datasets