aaire/owic25-embeddings

OWIC25 Embeddings 2026-11-04

Dense document embeddings of the whole OWIC25 web test collection (20,746,181 documents) from four open multilingual embedding models. One L2-normalised float16 vector per document and model (768-dimensional, 1,024 for harrier-oss-v1-0.6b), keyed by the owic25 document id. No document text.

The data is not stored on Hugging Face. It is served by the Open Web Index (92 GB in 1,500 parquet parts for the first three models; harrier-oss-v1-0.6b adds about 40 GB in 500 parts):

route	link
Open Web Index dashboard	https://openwebindex.eu/corpora/fe28515c-d91d-11f1-b6bc-421fe4d7e0dd
owilix	owilix remote pull all/internalID=fe28515c-d91d-11f1-b6bc-421fe4d7e0dd
OWIC25 itself (the text)	https://openwebindex.eu/corpora/b510baa6-ebd2-11f0-8c43-02a47ca5d9fd

The full datasheet, file format and checksums ship with the data (README.md, FORMAT.md, SHA256SUMS).

Models

directory	Hugging Face model @ revision	pooling	max tokens
arctic-embed-m-v2.0	Snowflake/snowflake-arctic-embed-m-v2.0 @ 95c2741	CLS	8,192
granite-embedding-311m-r2	ibm-granite/granite-embedding-311m-multilingual-r2 @ 4439955	CLS	8,192
embeddinggemma-2	google/embeddinggemma-2 @ 914f7f8	mean	8,192
harrier-oss-v1-0.6b	microsoft/harrier-oss-v1-0.6b @ f9b9dc8	last token	8,192

Queries must be encoded with the same model, revision, pooling and query prefix (each model's model.json); score with the dot product. arctic-embed-m-v2.0, granite-embedding-311m-r2 and embeddinggemma-2 support Matryoshka truncation (keep the first k dimensions, then L2-normalise); harrier-oss-v1-0.6b states none. harrier-oss-v1-0.6b was added to the release later (CHANGELOG.md in the data).

Layout

embeddings/<model>/float16/part-00000.parquet … part-00499.parquet   part i = owic25 metadata_<i>.parquet, row for row
embeddings/<model>/model.json, ranges.json
tools/derive_precisions.py                                         float16 -> int8 / int4 / int2 / binary

Columns: docid (owic25 id, sha256 of the URL), vec (fixed_size_list[768], [1024] for harrier-oss-v1-0.6b), input_sha256 (digest of the token ids encoded), n_tokens, truncated. The 66,214 documents without text keep a zero vector with n_tokens 0; leave them out of similarity search.

import pyarrow.parquet as pq, numpy as np
t = pq.read_table("embeddings/embeddinggemma-2/float16/part-00000.parquet")
X = t.column("vec").combine_chunks().flatten().to_numpy().reshape(len(t), 768)

Quality checks

Per model: 1,000 random documents re-encoded in mixed batches, and 61,326 documents compared with an independent earlier encoding (for the first three models every cosine ≥ 0.99999 and ≥ 0.9998; the bars every model passed are 0.9995 and 0.999; figures in each model.json → verify). The first three models took about 175 GPU hours on one RTX 4090; each model's time is in its model.json.

Licence and attribution

CC BY-SA 4.0, with a non-binding request that models trained on these embeddings stay openly available.

This work uses index files as part of the index partition created by the OpenWebSearch.eu project, funded by the European Union's Horizon Europe research and innovation programme under grant agreement No 101070014 (OpenWebSearch.EU).

Any content made available through the OWI is the property of its respective owners. Embeddings can partly be inverted to text, so the page owners' rights still apply. Takedown requests: https://openwebindex.eu/takedown.

Citation

Michael Granitzer, Laura Caspari, Michael Dinzinger, Kanishka Ghosh Dastidar, Saber Zerhoudi, Jelena Mitrovic. OWIC25 Embeddings 2026-11-04 (CITATION.cff ships with the data).