lance-format/vibe-msmarco-qwen-1024

VIBE msmarco-qwen-1024-normalized in Lance format

This repository is a re-packaging of the msmarco-qwen-1024-normalized dataset from VIBE: Vector Index Benchmark for Embeddings as two Lance datasets: the corpus vectors and the benchmark queries with their exact ground truth. Nothing else was changed: the vectors are byte-identical to the upstream HDF5 file and the row order is preserved.

No vector or scalar index is included. The datasets contain plain data only, so you can build whatever index you want to benchmark (IVF_PQ, IVF_RQ, HNSW, ...) on top of them.

Provenance

Upstream dataset	vector-index-bench/vibe, file msmarco-qwen-1024-normalized.hdf5
Paper	VIBE: Vector Index Benchmark for Embeddings (arXiv:2505.17810), code at vector-index-bench/vibe
Source text	MS MARCO passages
Embedding model	Qwen text embedding model, as used by VIBE (1024 dimensions)
Corpus rows	8,840,823
Queries	1,000, each with its exact 100 nearest neighbors
Dimensions	1024, float32
Normalization	all vectors are unit-norm
Distance metric	"any": cosine, inner product and euclidean give the same ranking on unit vectors
Upstream license	CC-BY-4.0

Contents

base.lance/      corpus, 8,840,823 rows, 9 fragments of <= 1,000,000 rows (~36 GB)
queries.lance/   1,000 queries with ground truth (< 10 MB)
README.md

base.lance

Column	Type	Meaning
id	int64	row index, 0 .. 8840822; equal to the row index in the upstream HDF5 train array
vector	fixed_size_list<float32>[1024]	the corpus embedding, byte-identical to train[id]

queries.lance

Column	Type	Meaning
query_id	int64	row index in the upstream HDF5 test array, 0 .. 999
vector	fixed_size_list<float32>[1024]	the query embedding, byte-identical to test[query_id]
neighbors	list<int64>	the 100 exact nearest corpus rows (base.lance id), best first, copied from the HDF5 neighbors array
scores	list<float32>	inner product (= cosine similarity) between the query and each neighbor, descending; computed from the vectors above

The row order of both datasets equals the upstream HDF5 row order, so base.lance row i is train[i] and queries.lance row j is test[j], and the neighbors column indexes directly into base.lance.

How to use

import lance
from huggingface_hub import snapshot_download

root = snapshot_download("lance-format/vibe-msmarco-qwen-1024", repo_type="dataset")

corpus = lance.dataset(f"{root}/base.lance")
queries = lance.dataset(f"{root}/queries.lance")
print(corpus.count_rows(), queries.count_rows(), corpus.list_indices())  # 8840823 1000 []

# Build your own index and run the benchmark queries
corpus.create_index("vector", index_type="IVF_PQ", metric="cosine", num_partitions=1024, num_sub_vectors=64)
q = queries.take([0]).column("vector")[0].values.to_numpy()
hits = corpus.to_table(nearest={"column": "vector", "q": q, "k": 100}, columns=["id"])

With LanceDB:

import lancedb
db = lancedb.connect(root)
tbl = db.open_table("base")
tbl.search(q).metric("cosine").limit(100).to_pandas()

Notes

• Written with pylance 12.0.0-beta.18, Lance file format version 2.0 (stable), one commit per dataset, no indices.
• Both datasets have exactly one version. No _indices directory exists.
• Validation performed before upload: row counts and dimension, 500 random rows plus the first and last row compared byte-for-byte with the upstream vectors, and a brute-force cosine top-10 over the whole corpus reproducing the stored ground truth for three queries.

License and attribution

The data is redistributed under the CC-BY-4.0 license of the upstream VIBE release. Please credit the VIBE authors when you use it:

@misc{vibe2025,
  title         = {VIBE: Vector Index Benchmark for Embeddings},
  author        = {Elias J{\"a}{\"a}saari and Ville Hyv{\"o}nen and Matteo Ceccarello and Teemu Roos and Martin Aum{\"u}ller},
  year          = {2025},
  eprint        = {2505.17810},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2505.17810}
}