maximerivest/openalex-denseon-embeddings

OpenAlex paper embeddings: 203.7 million English titles + abstracts, ready to search

One row per work in OpenAlex that has an English abstract: its OpenAlex id, DOI, title, date, type, venue, field, a 2D position on a map of science, and a 768-number embedding of its title and abstract made with lightonai/DenseOn. Papers about the same thing get nearby vectors, so you can search by meaning, find similar papers, cluster, or map all of science.

These are the vectors behind the search on sciencemadereadable.com and its free search API for people and AI agents.

Rows (embedded works)	203,725,202
…of which in the search index	125,130,657 publications, one per distinct text (in_search_index)
Size	297 GB in 502 Parquet files (0.74 GB each at most), split by publication year
Text embedded	title + " " + abstract, English only, first 512 tokens
Vector	768 × float16, L2-normalised (cosine similarity = dot product)
Metadata as of	OpenAlex snapshot 2026-09-23 + daily OpenAlex API additions to 2026-10-12
Data release	20261012T073823Z

Pick your path

You have	Do this
Nothing to download	Use the hosted API: curl https://ask.sciencemadereadable.com/coral+reef+recovery (docs)
A laptop	Query the metadata remotely with DuckDB (reads only the columns you ask for), or download one year
~20 GB of RAM, ~16 GB of disk	Download index/ and search all 125 million papers by meaning (below)
~350 GB of SSD	Download everything: exact re-scoring, your own index, clustering, maps

Look without downloading (DuckDB)

Parquet lets DuckDB fetch only the columns and row groups a query needs, straight from the Hub:

-- duckdb
SELECT year, count(*) AS papers
FROM 'hf://datasets/maximerivest/openalex-denseon-embeddings/data/**/*.parquet'
WHERE year BETWEEN 2015 AND 2026
GROUP BY year ORDER BY year;

Download part of it

Files are grouped by publication year (data/2025/, data/1999/, …, plus data/before-1900/, data/after-2026/, data/no-year/ and the daily additions in data/updates/):

pip install -U huggingface_hub
hf download maximerivest/openalex-denseon-embeddings --repo-type dataset --include "data/2025/*" --local-dir openalex-denseon
hf download maximerivest/openalex-denseon-embeddings --repo-type dataset --local-dir openalex-denseon          # everything, ~297 GB

Use the vectors

import numpy as np, pyarrow.parquet as pq
from sentence_transformers import SentenceTransformer

t = pq.read_table("openalex-denseon/data/2025/part-00000.parquet")
E = t["embedding"].combine_chunks().values.to_numpy().reshape(-1, 768)   # float16, one row per paper

model = SentenceTransformer("lightonai/DenseOn", revision="cb9947ebccb33862d24e3c7ca2edb25e51acd887")
q = model.encode("coral reef recovery after bleaching", prompt_name="query", normalize_embeddings=True)
scores = E.astype(np.float32) @ q
for i in np.argsort(-scores)[:5]:
    print(f"{scores[i]:.3f}", t["title"][i], "https://openalex.org/" + str(t["openalex_id"][i]))

Questions are encoded with prompt_name="query" ("query: " in front), papers with prompt_name="document" ("document: "). To put your own texts in the same space, encode title + " " + abstract with the document prompt: re-encoding papers on a CPU in float32 gives a cosine ≥ 0.999 with the stored vectors (we checked).

With 🤗 Datasets, stream rather than load: a plain load_dataset converts all 297 GB into its cache first.

from datasets import load_dataset
ds = load_dataset("maximerivest/openalex-denseon-embeddings", split="train", streaming=True)
print(next(iter(ds))["title"])

Search everything on your own machine

index/ holds the search index the live site uses for this release: FAISS IVF32768_HNSW32,RaBitQ over the 125,130,657 rows with in_search_index = true (14.1 GB), and index/work_ids.npy (row number → OpenAlex W-number, 1.6 GB). Its ids are the row column.

hf download maximerivest/openalex-denseon-embeddings --repo-type dataset --include "index/*" --include "*.py" --local-dir openalex-denseon
cd openalex-denseon
uv run search.py "coral reef recovery after bleaching"          # needs ~20 GB of RAM

What you downloaded	Quality of the top 10	Time per question
index/ only	~75% of an exact search's top 10 (approximate scores)	~5–30 ms, after a ~20 s load
index/ + all of data/ + uv run make_vectors.py (builds a 313 GB vectors.npy; use an SSD)	~99% (300 candidates re-scored exactly)	~10–50 ms
data/ for some years, no index	exact, over those years	~1 minute per recent year from an SSD (each is 15–40 GB)

(Measured on 200 paper-vector questions against a deeper exact re-scoring, release 20261012T073823Z; not a relevance judgement. The index was written by faiss-cpu==1.15.1, which search.py pins.)

The live site ranks a little differently: it also demotes likely junk records (spam, paratext), which this dataset does not include.

Columns

Column	Type	What it is
row	int64	Row number, 0 … 203,725,201. The index's ids. New versions add rows at the end (Versions)
openalex_id	string	W… id; https://openalex.org/<id> is the record, https://api.openalex.org/works/<id> its full metadata
doi	string	DOI without https://doi.org/, null if none
title	string	Title as embedded
publication_date, year	date, int16	From OpenAlex. A few are wrong in OpenAlex (years like 1007 or 2050)
type	string	OpenAlex type: article, book-chapter, preprint, review, paratext, erratum, …
venue	string	Journal, conference or repository
field, subfield, topic	string	OpenAlex primary topic and its subfield and field
cited_by_count	int32	Citations counted by OpenAlex at that time
is_retracted, is_open_access, license	bool, bool, string	From OpenAlex (license is the work's licence, e.g. cc-by)
in_search_index	bool	In the search index: a publication type (article, review, preprint, book, chapter, conference paper, thesis, report, letter, editorial, …) and the smallest W-number among works with byte-identical title + abstract
atlas_x, atlas_y	float32	Position (0–1) on the map of science: openTSNE layout of a reference sample, every paper placed from its 10 nearest reference papers. Atlas 20261009T190750-tsne-mode. Projections distort distances: use it to look, not to measure
text_hash	binary(16)	BLAKE2b-128 of the UTF-8 text embedded; equal hash = identical text
embedding	float16[768]	The vector

Every Parquet file also carries the model, revision, prompts and input format in its metadata.

How it was made

• Selection. All OpenAlex works whose language is en and whose abstract is not empty, with no filter on type, open access or completeness (203,725,202 rows: 203,486,035 from the OpenAlex snapshot of 2026-09-23, 239,167 new papers from the daily API run since).
• Text. ((title or "") + " " + abstract).strip(), abstract rebuilt from OpenAlex's inverted index (words put back in order). DenseOn reads the first 512 tokens.
• Model. lightonai/DenseOn at revision cb9947ebccb33862d24e3c7ca2edb25e51acd887 (149M parameters, Apache-2.0), sentence-transformers 5.3.0, bfloat16 on an RTX 3090, document: prompt, then L2-normalised in float32 and stored as float16.
• Checks. For every year folder, the title + abstract of a full shard of rows was hashed again from the OpenAlex records and compared with the hash stored next to each vector: all matched. Every row was joined to exactly one OpenAlex record.

Limits, said plainly

• Not all of science. Only works OpenAlex labels English and gives an abstract. Many papers (older ones, many books, much non-English research) have no abstract in OpenAlex, and its language labels make mistakes in both directions.
• Not only papers. 73,731,213 rows (36%) are not publications by the definition above (dataset 57.3M, other 11.8M, paratext 0.8M, software 0.7M, …), and 4,863,332 publications repeat another record's exact title and abstract. in_search_index removes both; type and text_hash let you choose differently.
• Metadata ages. Titles, citation counts, topics and retractions are as of 2026-09-23 (daily additions: as of the day they were fetched). OpenAlex re-ran its topic classifier after that snapshot, so recent additions carry newer topic labels than older rows.
• No abstracts here. Abstracts belong to their publishers, and this dataset does not redistribute them. Get them from OpenAlex (free snapshot or API) by openalex_id, or through our API.
• Search quality of the vectors is DenseOn's (BEIR nDCG@10 56.2 per its authors). We have not published a relevance evaluation of this release.

Versions

New versions add files for new papers and may refresh metadata columns; an existing row keeps its row number and vector. A change that would renumber rows or change vectors (a new model, a new way to build the text) will be a separate dataset, not a silent update. Each version is a tagged revision of this repository named after its data release; export.json describes it.

Licence

CC0 1.0: no rights reserved. Use, share and build on it for anything, commercial or not, without asking. Credit is appreciated, not required (see Cite).

The metadata comes from OpenAlex (CC0). DenseOn is Apache-2.0.

Cite

@misc{rivest_openalex_denseon_2026,
  author = {Rivest, Maxime},
  title  = {OpenAlex paper embeddings (DenseOn, English titles and abstracts)},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/maximerivest/openalex-denseon-embeddings}
}

Also cite OpenAlex (Priem, Piwowar & Orr 2022, arXiv:2205.01833) and DenseOn. Made by ScienceMadeReadable; code: srl-search, pipeline.