lance-format/laion2b-en-clip-vit-l14-10m

LAION-2B-en CLIP ViT-L/14 image embeddings, 10M x 768 (Lance format)

10,000,000 base vectors and 4,992 query vectors, 768-d float32, with exact cosine ground truth, packaged as Lance datasets. It is the dataset behind LanceDB's "10M real embeddings: IVF index comparison" benchmark (IVF_RQ 1/3/5 bit vs IVF_PQ vs IVF_SQ).

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, IVF_SQ, HNSW, ...) on top of them.

Provenance

Source	the CLIP ViT-L/14 image embeddings (img_emb) that LAION distributes with LAION-2B-en; the same embedding family is used by the SISAP 2023 indexing challenge (laion2B-en-clip768v2)
Rows	a 30M-row sample of LAION-2B-en: rows 0 .. 9,999,999 are the base, rows 10,000,000 .. 10,004,999 are the queries
Dimensions	768
Stored dtype	float32, an exact cast of the upstream float16 values (no normalization, no other change)
Distance metric	cosine
Ground truth	exhaustive float32 cosine search over all 10M base vectors (GPU), top 4,096 per query, validated against a CPU search

Of the 5,000 query rows, 8 were dropped because the tie group at the top-10 boundary extended past the 4,096 stored neighbors, leaving 4,992 queries. 104 of the remaining queries have exact ties at the top-10 boundary; see Recall with ties.

Only embeddings are included: no images, captions, URLs or other LAION metadata.

Contents

base.lance/      10,000,000 rows, 10 fragments of 1,000,000 rows (~29 GB)
queries.lance/   4,992 queries with ground truth (~250 MB)
README.md

base.lance

Column	Type	Meaning
vector	fixed_size_list<float32>[768]	the image embedding

The row position (0-based, in storage order) is the id used by the ground truth.

queries.lance

Column	Type	Meaning
query_id	uint64	the query's row in the 30M sample (10,000,000 ..)
vector	fixed_size_list<float32>[768]	the query embedding
neighbors	list<int64>	the 4,096 exact nearest base rows (0-based positions in base.lance), best first; equal distances are ordered by row
distances	list<float32>	the cosine distance (1 - cos) of each neighbor, ascending

How to use

import lance
from huggingface_hub import snapshot_download

root = snapshot_download("lance-format/laion2b-en-clip-vit-l14-10m", repo_type="dataset")

base = lance.dataset(f"{root}/base.lance")
queries = lance.dataset(f"{root}/queries.lance")
print(base.count_rows(), queries.count_rows(), base.list_indices())  # 10000000 4992 []

base.create_index("vector", index_type="IVF_RQ", metric="cosine", num_bits=5, target_partition_size=4096)
q = queries.take([0], columns=["vector", "neighbors"])
hits = base.to_table(
    nearest={"column": "vector", "q": q["vector"][0].values.to_numpy(), "k": 10, "nprobes": 64, "metric": "cosine"},
    columns=[],
    with_row_id=True,
)

_rowid values are not row positions in general; map them back to positions before comparing with neighbors (for this single-version dataset, position = fragment_id * 1,000,000 + (_rowid & 0xFFFFFFFF)).

Recall with ties

For recall@k, count a returned row as correct if it is in neighbors[:k] or if its distance equals distances[k - 1] (an exact tie at the boundary). Every tie group at the top-10 boundary is fully contained in the stored 4,096 neighbors.

Reference results

IVF with target_partition_size=4096 (2,441 partitions; RQ3/RQ5 reuse the RQ1 centroids), k=10, nprobes=64, no refine, index prewarmed, 32 concurrent client threads, median of 3 x 10,000 requests, pylance 12.0.0 on an AWS r8i.8xlarge (32 vCPU):

Index	Size GiB	Recall@10	QPS	Mean ms	P99 ms
IVF_RQ 1 bit	1.09	0.7654	1806	16.9	29.1
IVF_RQ 3 bit	2.95	0.9202	1790	17.1	29.3
IVF_RQ 5 bit	4.74	0.9623	1781	17.2	29.4
IVF_PQ 96x8	0.98	0.6922	1360	23.1	29.7
IVF_SQ 8 bit	7.23	0.9556	672	47.4	55.6

Notes

• Written with pylance 12.0.0, Lance file format 2.0 (readable by older pylance releases), one version per dataset, no indices.
• Integrity checks before upload: SHA-256 of the base vectors as little-endian float32 in row order is 9d7671f3d33f55814315b33880e401f87e1d2e94d2b4051a56d0310edbae076e, and of the 4,992 query vectors 7f6bf1fca76697bd246db7e3b7b37eb737e6998733b550346e0e3c09c705a094; a brute-force cosine top-10 over the published base reproduces the stored ground truth for sampled queries.

License and attribution

The embeddings are derived from LAION-2B-en, released by LAION under CC-BY-4.0; this repository redistributes them under the same license. Please credit LAION (Schuhmann et al., "LAION-5B: An open large-scale dataset for training next generation image-text models", NeurIPS 2022) and OpenAI for the CLIP ViT-L/14 model. The underlying images belong to their respective owners and are not included.