BEIR — unified schema

14 datasets of the BEIR benchmark as mteb defines it — every task except CQADupstack (twelve StackExchange sub-forums, left out by choice) — each republished as its own repo in one strict, identical schema, with a provenance.json that ties it to the exact mteb/* source revision the leaderboard scores. Where BEIR provides train/dev splits, the queries and qrels configs carry them too; the corpus is shared by all splits. This repo is just the index; the data lives in the linked repos.

task	repo	languages	domain	license	splits	queries	documents	qrels	= mteb
ArguAna	beir-arguana	eng	argument	cc-by-sa-4.0	test	1,406	8,674	1,406	✓
ClimateFEVER	beir-climate-fever	eng	fact-checking	cc-by-sa-4.0	test	1,535	5,416,593	4,681	✓
DBPedia	beir-dbpedia	eng	entity	cc-by-sa-3.0	test, dev	467	4,635,922	49,188	✓
FEVER	beir-fever	eng	fact-checking	cc-by-nc-sa-3.0	test, train, dev	123,142	5,416,568	156,101	✓
FiQA2018	beir-fiqa-2018	eng	finance	cc-by-sa-4.0	test, train, dev	6,648	57,638	17,110	✓
HotpotQA	beir-hotpotqa	eng	wikipedia	cc-by-sa-4.0	test, train, dev	97,852	5,233,329	195,704	✓
MSMARCO	beir-msmarco	eng	web	msr-la-nc	test, train	509,919	8,841,823	540,188	✓
NFCorpus	beir-nfcorpus	eng	biomedical	cc-by-sa-4.0	test, train, dev	3,237	3,633	134,294	✓
NQ	beir-nq	eng	wikipedia	cc-by-nc-sa-3.0	test	3,452	2,681,468	4,201	✓
QuoraRetrieval	beir-quora	eng	duplicate-question	cc-by-sa-4.0	test, dev	15,000	522,931	23,301	✓
SCIDOCS	beir-scidocs	eng	scientific	cc-by-sa-4.0	test	1,000	25,657	29,928	✓
SciFact	beir-scifact	eng	scientific	cc-by-nc-4.0	test, train	1,109	5,183	1,258	✓
Touche2020	beir-touche-2020	eng	argument	cc-by-sa-4.0	test	49	382,545	2,214	✓
TRECCOVID	beir-trec-covid	eng	biomedical	cord-19	test	50	171,332	66,336	✓

"= mteb" ✓ means the data is the data the benchmark scores, proven two ways: a streaming digest of everything mteb's loader sees — every canonically ordered corpus row, every query, every qrel — is identical on both sides for every dataset; and, wherever a deterministic encoder has non-zero metrics (all but the million-document corpora, where a random ranker scores 0 on both sides), running mteb on the task as shipped and on this build gives identical metrics to 1e-9. (Corpus order is canonicalised on both sides first: documents with identical text tie, and mteb breaks ties by corpus position.) Counts are the evaluated split.

The schema (every repo)

config	columns	split
queries	id: string, text: string	test (+ train/dev where the benchmark has them)
corpus	id: string, title: string, text: string	test — shared by all splits
qrels	query-id: string, corpus-id: string, score: int32	test (+ train/dev where the benchmark has them)

Multilingual repos prefix the configs (ar-queries, ar-corpus, ar-qrels, …). Rules: ids unique and non-empty; title always present ("" if none); every query in a split has ≥ 1 qrel in that split; no duplicate pairs; referential integrity to the corpus; qrels keep the source's graded and zero scores.

Use with mteb

import mteb, types
from mteb.abstasks.retrieval import AbsTaskRetrieval

task = mteb.get_task("ArguAna")
task.metadata = task.metadata.model_copy(update={"dataset": {"path": "Hyukkyu/beir-arguana", "revision": "main"}})
# a few tasks ship a custom load_data hardcoded to their source's split names; the generic loader reads this layout
if type(task).load_data is not AbsTaskRetrieval.load_data:
    task.load_data = types.MethodType(AbsTaskRetrieval.load_data, task)
mteb.evaluate(mteb.get_model("intfloat/multilingual-e5-small"), [task])

Provenance and licensing

Each repo's provenance.json records the source repo and revision, the sha256 of every source file, every change made (renames, casts, dropped columns, orphan/duplicate counts), and the sha256 of every output file. Licenses are per dataset and follow the most restrictive terms an authoritative source states where sources disagree; the index carries none of its own. Not affiliated with the benchmark or mteb maintainers. Built 2026-09-08 from mteb main.