JMTEB-lite: The Lightweight Version of JMTEB

JMTEB-lite is a lightweight version of JMTEB. It makes agile evaluation possible by reaching an average of 5x faster evaluation comparing with JMTEB. The result of JMTEB-lite is proved to be highly relevant with that of JMTEB, making it a faithful preview of JMTEB.

TL;DR

from datasets import load_dataset

dataset = load_dataset("sbintuitions/JMTEB-lite", name="<dataset_name>", split="<split>")

SPLITS = ("train", "validation", "test", "corpus")
JMTEB_LITE_DATASET_NAMES = (
    'livedoor_news',
    'mewsc16_ja',
    'sib200_japanese_clustering',
    'amazon_review_classification',
    'amazon_counterfactual_classification',
    'massive_intent_classification',
    'massive_scenario_classification',
    'japanese_sentiment_classification',
    'sib200_japanese_classification',
    'wrime_classification',
    'jsts',
    'jsick',
    'jaqket-query',
    'jaqket-corpus',  # lightweight
    'mrtydi-query',
    'mrtydi-corpus',  # lightweight
    'jagovfaqs_22k-query',
    'jagovfaqs_22k-corpus',
    'nlp_journal_title_abs-query',
    'nlp_journal_title_abs-corpus',
    'nlp_journal_title_intro-query',
    'nlp_journal_title_intro-corpus',
    'nlp_journal_abs_intro-query',
    'nlp_journal_abs_intro-corpus',
    'nlp_journal_abs_article-query',
    'nlp_journal_abs_article-corpus',
    'jacwir-retrieval-query',
    'jacwir-retrieval-corpus',  # lightweight
    'miracl-retrieval-query',
    'miracl-retrieval-corpus',  # lightweight
    'mldr-retrieval-query',
    'mldr-retrieval-corpus',
    'mintaka-retrieval-query',
    'mintaka-retrieval-corpus',
    'esci-query',
    'esci-corpus',
    'jqara-query',
    'jqara-corpus',
    'jacwir-reranking-query',  # lightweight
    'jacwir-reranking-corpus',  # lightweight
    'miracl-reranking-query',  # lightweight
    'miracl-reranking-corpus',  # lightweight
    'mldr-reranking-query',
    'mldr-reranking-corpus',
)

Introduction

We introduced JMTEB (Japanese Massive Text Embedding Benchmark), a comprehensive evaluation benchmark of Japanese text embedding models. However, the massive size of JMTEB makes evaluation slow and resource demanding. To address this, we now introduce JMTEB-lite, a lightweight version of JMTEB constructed by substaintially reducing corpus size in retrieval and reranking tasks. We have also verified that JMTEB-lite significantly accelerates evaluation while maintaining high fidelity to the full JMTEB.

We recommand to use JMTEB-lite to obtain the preview evaluation results in agile development, and use JMTEB for full and final evaluation.

JMTEB-lite is compatible with the evaluation script of JMTEB: https://github.com/sbintuitions/JMTEB.

Tasks and Datasets

Here is an overview of the tasks and datasets currently included in JMTEB-lite.

Note that only datasets in bold are lightweight, and the rest are exactly the same with the counterparts in JMTEB.

Task	Dataset	Train	Dev	Test	Document (Retrieval)
Clustering	Livedoor-News	5,163	1,106	1,107	-
	MewsC-16-ja	-	992	992	-
	SIB200 Japanese Clustering	701	99	204	-
Classification	AmazonCounterfactualClassification	5,600	466	934	-
	AmazonReviewClassification	200,000	5,000	5,000	-
	MassiveIntentClassification	11,514	2,033	2,974	-
	MassiveScenarioClassification	11,514	2,033	2,974	-
	Japanese Sentiment Classification	9,831	1,677	2,552	-
	SIB200 Japanese Classification	701	99	204	-
	WRIME Classification	30,000	2,500	2,500	-
STS	JSTS	12,451	-	1,457	-
	JSICK	5,956	1,985	1,986	-
Retrieval	JAQKET	13,061	995	997	65,802
	Mr.TyDi-ja	3,697	928	720	93,382
	NLP Journal title-abs	-	127	510	637
	NLP Journal title-intro	-	127	510	637
	NLP Journal abs-intro	-	127	510	637
	NLP Journal abs-abstract	-	127	510	637
	JaGovFaqs-22k	15,955	3,419	3,420	22,794
	JaCWIR-Retrieval	-	1,000	4,000	302,638
	MIRACL-Retrieval	2,433	1,044	860	105,064
	MLDR-Retrieval	2,262	200	200	10,000
	Mintaka-Retrieval	-	2,313[^1]	2,313	2,313
Reranking	Esci	10,141	1,790	4,206	149,999
	JaCWIR-Reranking	-	1,000	4,000	188,033
	JQaRA	498	1,737	1,667	172,897
	MIRACL-Reranking	2,433	1,044	860	37,124
	MLDR-Reranking	2,262	200	200	5,339

[^1]: To keep consistent with MTEB where Mintaka-Retrieval doesn't have a validation set, we set our validation set the same as the test set.

Construction Process

For the 4 retrieval datasets (JAQKET, Mr.TyDi, JaCWIR-Retrieval, MIRACL-Retrieval), we use 5 highly performant models to predict hard negative documents for each query (the query's most 50 semantically similar documents in the corpus), and merge these hard negatives along with golden documents.

For the 2 reranking datasets (JQaRA, JaCWIR-Reranking), we use 5 highly performant models to rerank the documents for each query, and retain top-50 hard negative documents for each query. Then we merge these hard negatives with golden documents.

For the rest, they are kept exactly the same with their counterparts in JMTEB.

Reference

@misc{jmteb_lite,
    author = {Li, Shengzhe and Ohagi, Masaya and Ri, Ryokan and Fukuchi, Akihiko and Shibata, Tomohide and Kawahara, Daisuke},
    title = {{J}{M}{T}{E}{B}-lite: {T}he {L}ightweight {V}ersion of {JMTEB}},
    howpublished = {\url{https://huggingface.co/datasets/sbintuitions/JMTEB-lite}},
    year = {2025},
}

@techreport{jmteb_lite,
    author = {Li, Shengzhe and Ohagi, Masaya and Ri, Ryokan and Fukuchi, Akihiko and Shibata, Tomohide and Kawahara, Daisuke},
    title = {{JMTEB} and {JMTEB}-lite: {J}apanese {M}assive {T}ext {E}mbedding {B}enchmark and {I}ts {L}ightweight {V}ersion},
    institution = {SB Intuitions / Waseda University},
    number = {IPSJ SIG Technical Reports Vol.2025-NL-265 No.3},
    year = {2025},
    month = {09},
}

License

Regarding the license information of datasets, please refer to the individual datasets.

Our code is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License.