vector-index-bench/vibe

This repository contains the datasets that are meant to be used with VIBE (Vector Index Benchmark for Embeddings):

https://github.com/vector-index-bench/vibe

The datasets can be downloaded manually from this repository, but the benchmark framework also downloads them automatically.

Datasets

Name	Type	n	d	Distance
agnews-mxbai-1024-euclidean	Text	769,382	1024	euclidean
arxiv-nomic-768-normalized	Text	1,344,643	768	any
gooaq-distilroberta-768-normalized	Text	1,475,024	768	any
imagenet-clip-512-normalized	Image	1,281,167	512	any
landmark-nomic-768-normalized	Image	760,757	768	any
yahoo-minilm-384-normalized	Text	677,305	384	any
celeba-resnet-2048-cosine	Image	201,599	2048	cosine
ccnews-nomic-768-normalized	Text	495,328	768	any
codesearchnet-jina-768-cosine	Code	1,374,067	768	cosine
glove-200-cosine	Word	1,192,514	200	cosine
landmark-dino-768-cosine	Image	760,757	768	cosine
simplewiki-openai-3072-normalized	Text	260,372	3072	any
coco-nomic-768-normalized	Text-to-Image	282,360	768	any
imagenet-align-640-normalized	Text-to-Image	1,281,167	640	any
laion-clip-512-normalized	Text-to-Image	1,000,448	512	any
yandex-200-cosine	Text-to-Image	1,000,000	200	cosine
yi-128-ip	Attention	187,843	128	IP
llama-128-ip	Attention	256,921	128	IP

Credit

The glove-200-cosine dataset uses embeddings from Glove (released under PDDL 1.0): https://nlp.stanford.edu/projects/glove/

The laion-clip-512-normalized dataset uses a subset of embeddings from LAION-400M (released under CC-BY 4.0): https://laion.ai/blog/laion-400-open-dataset/

The yandex-200-cosine dataset uses a subset of embeddings from Yandex Text2Image (released under CC-BY 4.0): https://big-ann-benchmarks.com/neurips23.html

Dataset structure

Each dataset is distributed as an HDF5 file.

The HDF5 files contain the following attributes:

• dimension: The dimensionality of the data.
• distance: The distance metric to use.
• point_type: The precision of the vectors, one of "float", "uint8", or "binary".

The HDF5 files contain the following HDF5 datasets:

• train: numpy array of size (n_corpus, dim) containing the embeddings used to build the vector index
• test: numpy array of size (n_test, dim) containing the test query embeddings
• neighbors: numpy array of size (n_test, 100) containing the IDs of the true 100 k-nn of each test query
• distances: numpy array of size (n_test, 100) containing the distances of the true 100 k-nn of each test query
• avg_distances: numpy array of size n_test containing the average distance from each test query to the corpus points

Additionally, the HDF5 files of OOD datasets contain the following HDF5 datasets:

• learn: numpy array of size (n_learn, dim) containing a larger sample from the query distribution
• learn_neighbors: numpy array of size (n_learn, 100) containing the true 100 k-nn (from the corpus) for each point in learn