Training — unified schema

Retrieval training sets — pairs and triplets from the sentence-transformers collection, DPR, and the training splits of three BEIR tasks — each republished in the same strict schema as the RTEB and BEIR collections, plus two training tables: hard-negatives and teacher-scores. Both ship empty except where a source itself labels negatives (AllNLI contradictions, Quora non-duplicates, source = "dataset"): negatives are mined and teacher scores computed separately. Ids are sha1(text)[:20] where the source has none; identical texts collapse to one document. The proof here is pair recovery — every source pair and negative is reconstructible from the build with byte-equal text (provenance.json → recovery). This repo is just the index; the data lives in the linked repos.

The 15 training sources

They live in 13 repositories: nine have one of their own, three Wikipedia QA sets share the DPR corpus through prefixed configs, and three are the train splits of the BEIR copies (their corpus is the test split, shared by every split). Counts are the train split of each source.

source	repo	configs	domain	license	queries	qrels	documents
AG News	train-agnews	queries …	news title → description	unspecified	715,929	794,921	770,382
AllNLI	train-all-nli	queries …	NLI, symmetric	cc-by-sa-4.0	282,442	331,555	620,710
CC-News	train-ccnews	queries …	news title → article	other	489,515	538,526	496,328
ELI5	train-eli5	queries …	long-form QA (Reddit)	unspecified	323,136	325,473	325,293
NPR	train-npr	queries …	news title → body	unspecified	568,581	583,746	582,524
PAQ	train-paq	queries …	synthetic Wikipedia QA	cc-by-sa-3.0	64,371,441	64,371,441	9,004,577
Quora duplicates	train-quora-duplicates	queries …	duplicate questions, symmetric	other	86,169	149,263	123,703
SearchQA	train-searchqa	queries …	Jeopardy QA / web snippets	unspecified	117,220	581,652	578,015
SimpleWiki	train-simplewiki	queries …	Wikipedia simplification	cc-by-sa-3.0	101,784	102,078	101,994
Natural Questions	train-dpr-wikipedia	nq-queries …	Wikipedia QA (DPR)	cc-by-sa-3.0	58,622	439,345	21,015,324 †
TriviaQA	train-dpr-wikipedia	trivia-queries …	trivia QA (DPR)	cc-by-sa-3.0	60,296	730,095	21,015,324 †
SQuAD	train-dpr-wikipedia	squad-queries …	Wikipedia RC (DPR)	cc-by-sa-3.0	70,038	369,436	21,015,324 †
MS MARCO	beir-msmarco	queries (train) …	web search	msr-la-nc	502,939	532,751	8,841,823
HotpotQA	beir-hotpotqa	queries (train) …	multi-hop Wikipedia QA	cc-by-sa-4.0	85,000	170,000	5,233,329
FEVER	beir-fever	queries (train) …	fact verification	cc-by-nc-sa-3.0	109,810	140,085	5,416,568

† one shared corpus: the three DPR sets are three query sets over the same 21M Wikipedia passages, so they are counted once in the total.

Totals: 67.9M training queries and 70.2M qrels over 53.1M documents. PAQ alone is 64.4M of those queries; a mixture typically caps the large sources (100k queries each is a common choice) and keeps MS MARCO and the DPR sets whole.

Where each source came from

• Nine own repositories — the sentence-transformers pair sets, one repo each.
• train-dpr-wikipedia — Natural Questions, TriviaQA and SQuAD as DPR built them, over one 21M-passage Wikipedia corpus. Each is a config prefix (nq-, trivia-, squad-) on queries, qrels, hard-negatives and teacher-scores; corpus has no prefix because all three share it. These are the DPR builds, not the sentence-transformers pair files: fewer queries, but a real corpus to retrieve against.
• Three BEIR copies — MS MARCO, HotpotQA and FEVER have training splits in BEIR, so their training data lives in the BEIR — unified schema repositories rather than in a separate train-* one. queries/qrels carry train (and dev, test where the benchmark has them); corpus is the test split and is shared by every split; hard-negatives and teacher-scores exist for train.

Not included: Yahoo Answers (Webscope non-commercial), WikiAnswers duplicates, and WikiHow (no pair set on the Hub).

The schema (every repo)

config	columns	split
queries	id: string, text: string	test (+ train/dev where the benchmark has them)
hard-negatives	query-id, corpus-id, rank: int32, source	training repos; one row per mined negative
teacher-scores	query-id, corpus-id, teacher, score: float32	training repos; one row per scored pair, positives included
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)

Repos holding several query sets over one corpus prefix the configs (nq-queries, nq-qrels, …); multilingual repos do the same by language (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.

datasets cannot return a 0-example split, so while hard-negatives and teacher-scores are empty, read them directly with pyarrow/polars/pandas; the schema is declared in the file and in each repo's configs:.

Load it

from datasets import load_dataset

# a source with its own repo
q = load_dataset("Hyukkyu/train-agnews", "queries", split="train")

# one of the three DPR query sets, over the shared corpus
q = load_dataset("Hyukkyu/train-dpr-wikipedia", "nq-queries", split="train")
c = load_dataset("Hyukkyu/train-dpr-wikipedia", "corpus", split="train")

# a BEIR training split
q = load_dataset("Hyukkyu/beir-msmarco", "queries", split="train")
c = load_dataset("Hyukkyu/beir-msmarco", "corpus", split="test")   # one corpus for every split

Use with mteb

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

task = mteb.get_task("AGNews")
task.metadata = task.metadata.model_copy(update={"dataset": {"path": "Hyukkyu/train-agnews", "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; index updated 2026-09-09 to cover all 15 sources.