xquantize/biomed-retrieval-jax-pmc-10k

PMC open-access abstracts (10,000) with MedCPT embeddings

Research data, not for clinical use. These abstracts and embeddings support a retrieval demo and must not be used to inform medical decisions.

Titles and abstracts of 10,000 articles released in PubMed Central from 2026-01-01 to 2026-01-14 under CC0 or CC BY licences, with embeddings from NCBI's MedCPT article encoder, computed in JAX and verified against Hugging Face Transformers. Built with biomed-retrieval-jax, which searches them in its biomed-search command and its demo.

Files

• documents.jsonl: one article per line, with pmcid, pmid, doi, title, abstract (structured abstracts' sections labelled as "LABEL: text"), journal, year, citation, license ("CC0" or "CC BY"), publication_types and authors ("Surname Initials", in PubMed's order).
• embeddings.npy: a float32 array of shape (10,000, 768); row i is the embedding of line i of documents.jsonl, from ncbi/MedCPT-Article-Encoder at revision d05a736da4bb84ee4057b7f7999485be6ed85465, on (title, abstract) pairs, with [CLS] pooling, truncated at 512 tokens.
• manifest.json: how the corpus and embeddings were built: the query, dates and seed, every exclusion, the exact model revisions and the commits.

How the articles were selected

Candidates were the 26,144 articles released in PubMed Central from 2026-01-01 to 2026-01-14 with a CC0 or CC BY licence and no embargo, found with NCBI's E-utilities. In a seeded random order, each article's licence and retraction status were checked in its metadata on PMC's Cloud Service, and its title, abstract, authors and publication types were taken from PubMed, until 10,000 articles passed. An article was kept only if it is indexed in PubMed, has an abstract, and is not retracted, a retraction notice, an erratum or an expression of concern. Of the articles examined, these were excluded: 4,528 not indexed in PubMed (no PMID); 364 with no abstract; 158 of publication type Published Erratum; 69 of publication type Retraction Notice; 5 retracted articles; 4 of publication type Expression of Concern. In a sample of 145 of the articles not indexed in PubMed, 136 came from one journal's supplement of meeting abstracts.

Licences and attribution

Each article keeps its own licence, given in its license field: 9,963 are licensed CC BY and 37 CC0. Each record includes the article's authors, title, journal citation and DOI; if you reuse the texts, please credit their authors. The texts are taken from the articles' PubMed records. The code that built this dataset is licensed Apache-2.0. MedCPT is released by NCBI under a public-domain notice; please cite Jin et al., MedCPT, Bioinformatics 39(11), 2023.

Use

import json

import numpy as np
from huggingface_hub import snapshot_download

path = snapshot_download("xquantize/biomed-retrieval-jax-pmc-10k", repo_type="dataset")
documents = [json.loads(line) for line in open(f"{path}/documents.jsonl")]
embeddings = np.load(f"{path}/embeddings.npy")

To search, embed a query with ncbi/MedCPT-Query-Encoder (revision d83a36cc6b8e3a5c5e9d9d6ba156808c1643dcbc; at most 64 tokens, [CLS] pooling) and score the documents by dot product, as biomed-search does.

Limitations

• Two weeks of publications: many questions have no directly relevant article here.
• Only articles indexed in PubMed; retractions as known when the corpus was built.
• 166 articles are preprints, which have not been peer-reviewed: their publication_types include "Preprint".
• Embeddings are of the (title, abstract) pair truncated at 512 tokens; the texts are complete.
• The largest journals are Cureus (598 articles) and PloS one (335).