NYSgpt/biomed-eval

BioMed Eval

A frozen retrieval benchmark over biomedical scientific literature (433,449 papers).

BioMed Collection | BioMed Encoder | BioMed Reranker

The benchmark behind every number on the BioMed Encoder and BioMed Reranker cards: the queries, the gold labels, and the per-query results of every arm we scored — so anyone can check a published number, or run a paired significance test against their own model without rescoring ours.

• ✅ Frozen before any training, split by paper. 25% of papers held out by a declared hash rule; no benchmark gold paper appears anywhere in training.
• ✅ Every baseline measured here, not quoted. Same corpus, same queries, same metric code, each model with the similarity function its authors published it for.
• ✅ Per-query hit/miss for every arm. Paired exact McNemar and bootstrap CIs are reproducible from the files in results/.
• ✅ Keys-only. Queries and DOIs, not third-party text — every gold row carries the public-API URL to re-fetch its title and abstract.

---

Contents

file	rows	what
queries.jsonl	2,000	{qid, family, text, server, pub_year, published, abs_len} — the query string plus the segment keys used for per-segment reporting
qrels.tsv	2,000	qid, slug, doi, server, details_url — one gold paper per query; details_url returns its metadata and abstract from the server's public API
manifest.json	1	freeze date, split rule, embedding recipe, leak audit, provenance, limitations
subsample-manifest.json	1	how these 2,000 were drawn from the frozen 10,000, with the parent file's sha256
results/*.json	12 arms	metrics per family and per segment (server, era, published, abstract-length band)
results/*.perquery.json	9 arms	per-query rank of the gold paper, so any two arms can be compared with a paired test

Query families

family	queries	source
MESH	1,241	NLM MeSH descriptors assigned by indexers to the paper's published version — expert subject headings, the professional-search case
AUTHORKW	759	the authors' own keywords, as submitted

---

Performance

MeSH concept queries (n = 1,241), retrieved against all 433,449 papers, top-50.

Rank	Arm	R@1	R@10	nDCG@10	file
1	BioMed Encoder + BioMed Reranker	0.306	0.549	0.419	biomed-encoder+biomed-reranker
2	RRF(FTS + BioMed Encoder)	0.272	0.533	0.394	rrf_biomed-encoder+fts
3	ncbi/MedCPT + MedCPT-Cross-Encoder	0.291	0.525	0.405	medcpt+medcpt_ce
4	BioMed Encoder (retrieval only)	0.269	0.524	0.388	biomed-encoder
5	ncbi/MedCPT	0.235	0.457	0.337	medcpt
6	allenai/specter2_base	0.153	0.351	0.247	specter2
7	BAAI/bge-m3 (stock)	0.118	0.286	0.194	bge-m3
8	Postgres FTS	0.047	0.081	0.063	fts

Also in results/: BioMed Reranker over MedCPT's candidate pool (medcpt+biomed-reranker, 0.544 R@10), and three-way fusions.

Statistics on the cards — paired exact McNemar, 5,000-sample bootstrap CIs — are computed from the perquery files. To test your own model against any arm here, score the same 2,000 queries, record the rank of each gold paper, and pair by qid.

---

Details

Property
Corpus	bioRxiv (345,882) + medRxiv (87,567) preprints, harvested 2026-08-19 from the servers' public APIs
Unit of record	one paper (max(version) per (server, doi)); slug is the bare DOI suffix
Indexed text	`left(trim(title
Split	25% of papers held out by `md5("holdout:"
Selection	deterministic: lowest 10,000 by md5("eval:" + family + ":" + key_number), then lowest 2,000 by md5("scoresample:" + slug + ":" + query)
Metrics	R@1, R@5, R@10, R@50, MRR, nDCG@10 — per family and per segment
Frozen	2026-08-19T08:30Z

Leak audit at freeze time: 0 of 2,000 queries appear verbatim inside their own document.

---

How to use

import json, csv, requests

queries = [json.loads(l) for l in open("queries.jsonl")]
gold = {r["qid"]: r for r in csv.DictReader(open("qrels.tsv"), delimiter="\t")}

# Re-fetch a gold paper's title + abstract from the server's public API
r = requests.get(gold["MESH-000001"]["details_url"]).json()["collection"][-1]
print(r["title"]); print(r["abstract"])

# Paired comparison against a published arm
ours = {...}  # qid -> rank of the gold paper in your top-50 (or None)
theirs = json.load(open("results/biomed-encoder.perquery.json"))

Text for the full corpus is available the same way, one call per DOI, or in bulk from the servers' own API (https://api.biorxiv.org/details/biorxiv/{from}/{to}/{cursor}).

---

Licensing

Preprint titles and abstracts are distributed by the servers under per-paper licences, and the corpus is not uniformly permissive — a third of it is CC-BY-NC-ND and more than a quarter is all-rights-reserved. This release therefore ships no third-party text: queries (MeSH descriptors and author keywords), gold DOIs, and our own measurements. Those are ours to license and are released CC-BY-4.0. The text behind each DOI stays with its authors, under their licence, at its source.

---

📬 Contact

Questions, results, or a model to add to the table? Open a discussion in the Community tab.

Citation

@misc{biomedeval2026,
  title  = {BioMed Eval: a frozen retrieval benchmark over the complete bioRxiv and medRxiv corpus},
  author = {NYSgpt},
  year   = {2026},
  url    = {https://huggingface.co/datasets/NYSgpt/biomed-eval}
}