mteb/MMTEB-Multilingual-v2

MTEB(Multilingual, v2)

MTEB(Multilingual, v2)

An MTEB benchmark

Massive Text Embedding Benchmark

The large-scale multilingual expansion of MTEB known as MMTEB, driven mainly by highly-curated community contributions covering 250+ languages.

Tasks

Task	Type	Description
BornholmBitextMining	semantic-similarity	Danish Bornholmsk Parallel Corpus. Bornholmsk is a Danish dialect spoken on the island of Bornholm, Denmark. Historically it is a part of east Danish which was also spoken in Scania and Halland, Sweden.
BibleNLPBitextMining	semantic-similarity	Partial Bible translations in 829 languages, aligned by verse.
BUCC.v2	semantic-similarity	BUCC bitext mining dataset train split, gold set only.
DiaBlaBitextMining	semantic-similarity	English-French Parallel Corpus. DiaBLa is an English-French dataset for the evaluation of Machine Translation (MT) for informal, written bilingual dialogue.
FloresBitextMining	semantic-similarity	FLORES is a benchmark dataset for machine translation between English and low-resource languages.
IN22GenBitextMining	semantic-similarity	IN22-Gen is a n-way parallel general-purpose multi-domain benchmark dataset for machine translation spanning English and 22 Indic languages.
IndicGenBenchFloresBitextMining	semantic-similarity	Flores-IN dataset is an extension of Flores dataset released as a part of the IndicGenBench by Google
NollySentiBitextMining	semantic-similarity	NollySenti is Nollywood movie reviews for five languages widely spoken in Nigeria (English, Hausa, Igbo, Nigerian-Pidgin, and Yoruba.
NorwegianCourtsBitextMining	semantic-similarity	Nynorsk and Bokmål parallel corpus from Norwegian courts. Norwegian courts have two standardised written languages. Bokmål is a variant closer to Danish, while Nynorsk was created to resemble regional dialects of Norwegian.
NTREXBitextMining	semantic-similarity	NTREX is a News Test References dataset for Machine Translation Evaluation, covering translation from English into 128 languages. We select language pairs according to the M2M-100 language grouping strategy, resulting in 1916 directions.
NusaTranslationBitextMining	semantic-similarity	NusaTranslation is a parallel dataset for machine translation on 11 Indonesia languages and English.
NusaXBitextMining	semantic-similarity	NusaX is a parallel dataset for machine translation and sentiment analysis on 11 Indonesia languages and English.
Tatoeba	semantic-similarity	1,000 English-aligned sentence pairs for each language based on the Tatoeba corpus
BulgarianStoreReviewSentimentClassfication	classification	Bulgarian online store review dataset for sentiment classification.
CzechProductReviewSentimentClassification	classification	User reviews of products on Czech e-shop Mall.cz with 3 sentiment classes (positive, neutral, negative)
GreekLegalCodeClassification	classification	Greek Legal Code Dataset for Classification. (subset = chapter)
DBpediaClassification	classification	DBpedia14 is a dataset of English texts from Wikipedia articles, categorized into 14 non-overlapping classes based on their DBpedia ontology.
FinancialPhrasebankClassification	classification	Polar sentiment dataset of sentences from financial news, categorized by sentiment into positive, negative, or neutral.
PoemSentimentClassification	classification	Poem Sentiment is a sentiment dataset of poem verses from Project Gutenberg.
ToxicConversationsClassification	classification	Collection of comments from the Civil Comments platform together with annotations if the comment is toxic or not.
TweetTopicSingleClassification	classification	Topic classification dataset on Twitter with 6 labels. Each instance of TweetTopic comes with a timestamp which distributes from September 2019 to August 2021. Tweets were preprocessed before the annotation to normalize some artifacts, converting URLs into a special token {{URL}} and non-verified usernames into {{USERNAME}}. For verified usernames, we replace its display name (or account name) with symbols {@}.
EstonianValenceClassification	classification	Dataset containing annotated Estonian news data from the Postimees and Õhtuleht newspapers.
FilipinoShopeeReviewsClassification	classification	The Shopee reviews tl 15 dataset is constructed by randomly taking 2100 training samples and 450 samples for testing and validation for each review star from 1 to 5. In total, there are 10500 training samples and 2250 each in validation and testing samples.
GujaratiNewsClassification	classification	A Gujarati dataset for 3-class classification of Gujarati news articles
SentimentAnalysisHindi	classification	Hindi Sentiment Analysis Dataset
IndonesianIdClickbaitClassification	classification	The CLICK-ID dataset is a collection of Indonesian news headlines that was collected from 12 local online news publishers.
ItaCaseholdClassification	classification	An Italian Dataset consisting of 1101 pairs of judgments and their official holdings between the years 2019 and 2022 from the archives of Italian Administrative Justice categorized with 64 subjects.
KorSarcasmClassification	classification	The Korean Sarcasm Dataset was created to detect sarcasm in text, which can significantly alter the original meaning of a sentence. 9319 tweets were collected from Twitter and labeled for sarcasm or not_sarcasm. These tweets were gathered by querying for: irony sarcastic, and sarcasm. The dataset was created by gathering HTML data from Twitter. Queries for hashtags that include sarcasm and variants of it were used to return tweets. It was preprocessed by removing the keyword hashtag, urls and mentions of the user to preserve anonymity.
KurdishSentimentClassification	classification	Kurdish Sentiment Dataset
MacedonianTweetSentimentClassification	classification	An Macedonian dataset for tweet sentiment classification.
AfriSentiClassification	classification	AfriSenti is the largest sentiment analysis dataset for under-represented African languages.
AmazonCounterfactualClassification	classification	A collection of Amazon customer reviews annotated for counterfactual detection pair classification.
CataloniaTweetClassification	classification	This dataset contains two corpora in Spanish and Catalan that consist of annotated Twitter messages for automatic stance detection. The data was collected over 12 days during February and March of 2019 from tweets posted in Barcelona, and during September of 2018 from tweets posted in the town of Terrassa, Catalonia. Each corpus is annotated with three classes: AGAINST, FAVOR and NEUTRAL, which express the stance towards the target - independence of Catalonia.
CyrillicTurkicLangClassification	classification	Cyrillic dataset of 8 Turkic languages spoken in Russia and former USSR
IndicLangClassification	classification	A language identification test set for native-script as well as Romanized text which spans 22 Indic languages.
MasakhaNEWSClassification	classification	MasakhaNEWS is the largest publicly available dataset for news topic classification in 16 languages widely spoken in Africa. The train/validation/test sets are available for all the 16 languages.
MassiveIntentClassification	classification	MASSIVE: A 1M-Example Multilingual Natural Language Understanding Dataset with 51 Typologically-Diverse Languages
MultiHateClassification	classification	Hate speech detection dataset with binary (hateful vs non-hateful) labels. Includes 25+ distinct types of hate and challenging non-hate, and 11 languages.
NordicLangClassification	classification	A dataset for Nordic language identification.
NusaParagraphEmotionClassification	classification	NusaParagraphEmotionClassification is a multi-class emotion classification on 10 Indonesian languages from the NusaParagraph dataset.
NusaX-senti	classification	NusaX is a high-quality multilingual parallel corpus that covers 12 languages, Indonesian, English, and 10 Indonesian local languages, namely Acehnese, Balinese, Banjarese, Buginese, Madurese, Minangkabau, Javanese, Ngaju, Sundanese, and Toba Batak. NusaX-Senti is a 3-labels (positive, neutral, negative) sentiment analysis dataset for 10 Indonesian local languages + Indonesian and English.
ScalaClassification	classification	ScaLa a linguistic acceptability dataset for the mainland Scandinavian languages automatically constructed from dependency annotations in Universal Dependencies Treebanks. Published as part of 'ScandEval: A Benchmark for Scandinavian Natural Language Processing'
SwissJudgementClassification	classification	Multilingual, diachronic dataset of Swiss Federal Supreme Court cases annotated with the respective binarized judgment outcome (approval/dismissal)
NepaliNewsClassification	classification	A Nepali dataset for 7500 news articles
OdiaNewsClassification	classification	A Odia dataset for 3-class classification of Odia news articles
PunjabiNewsClassification	classification	A Punjabi dataset for 2-class classification of Punjabi news articles
PolEmo2.0-OUT	classification	A collection of Polish online reviews from four domains: medicine, hotels, products and school. The PolEmo2.0-OUT task is to predict the sentiment of out-of-domain (products and school) reviews using models train on reviews from medicine and hotels domains.
PAC	classification	Polish Paraphrase Corpus
SinhalaNewsClassification	classification	This file contains news texts (sentences) belonging to 5 different news categories (political, business, technology, sports and Entertainment). The original dataset was released by Nisansa de Silva (Sinhala Text Classification: Observations from the Perspective of a Resource Poor Language, 2015).
CSFDSKMovieReviewSentimentClassification	classification	The dataset contains 30k user reviews from csfd.cz in Slovak.
SiswatiNewsClassification	classification	Siswati News Classification Dataset
SlovakMovieReviewSentimentClassification	classification	User reviews of movies on the CSFD movie database, with 2 sentiment classes (positive, negative)
SwahiliNewsClassification	classification	Dataset for Swahili News Classification, categorized with 6 domains (Local News (Kitaifa), International News (Kimataifa), Finance News (Uchumi), Health News (Afya), Sports News (Michezo), and Entertainment News (Burudani)). Building and Optimizing Swahili Language Models: Techniques, Embeddings, and Datasets
DalajClassification	classification	A Swedish dataset for linguistic acceptability. Available as a part of Superlim.
TswanaNewsClassification	classification	Tswana News Classification Dataset
IsiZuluNewsClassification	classification	isiZulu News Classification Dataset
WikiCitiesClustering	clustering	Clustering of Wikipedia articles of cities by country from https://huggingface.co/datasets/wikipedia. Test set includes 126 countries, and a total of 3531 cities.
MasakhaNEWSClusteringS2S	clustering	Clustering of news article headlines from MasakhaNEWS dataset. Clustering of 10 sets on the news article label.
RomaniBibleClustering	clustering	Clustering verses from the Bible in Kalderash Romani by book.
ArXivHierarchicalClusteringP2P	clustering	Clustering of titles+abstract from arxiv. Clustering of 30 sets, either on the main or secondary category
ArXivHierarchicalClusteringS2S	clustering	Clustering of titles from arxiv. Clustering of 30 sets, either on the main or secondary category
BigPatentClustering.v2	clustering	Clustering of documents from the Big Patent dataset. Test set only includes documents belonging to a single category, with a total of 9 categories.
BiorxivClusteringP2P.v2	clustering	Clustering of titles+abstract from biorxiv across 26 categories.
MedrxivClusteringP2P.v2	clustering	Clustering of titles+abstract from medrxiv across 51 categories.
StackExchangeClustering.v2	clustering	Clustering of titles from 121 stackexchanges. Clustering of 25 sets, each with 10-50 classes, and each class with 100 - 1000 sentences.
AlloProfClusteringS2S.v2	clustering	Clustering of document titles from Allo Prof dataset. Clustering of 10 sets on the document topic.
HALClusteringS2S.v2	clustering	Clustering of titles from HAL (https://huggingface.co/datasets/lyon-nlp/clustering-hal-s2s)
SIB200ClusteringS2S	clustering	SIB-200 is the largest publicly available topic classification dataset based on Flores-200 covering 205 languages and dialects annotated. The dataset is annotated in English for the topics, science/technology, travel, politics, sports, health, entertainment, and geography. The labels are then transferred to the other languages in Flores-200 which are human-translated.
WikiClusteringP2P.v2	clustering	Clustering of wikipedia articles inspired by BlubrbsClusteringP2P. Labels are taken from top-level categories of the respective languages (e.g., https://lv.wikipedia.org/wiki/Kategorija:Pamatkategorijas).
PlscClusteringP2P.v2	clustering	Clustering of Polish article titles+abstracts from Library of Science (https://bibliotekanauki.pl/), either on the scientific field or discipline.
SwednClusteringP2P	clustering	The SWE-DN corpus is based on 1,963,576 news articles from the Swedish newspaper Dagens Nyheter (DN) during the years 2000--2020. The articles are filtered to resemble the CNN/DailyMail dataset both regarding textual structure. This dataset uses the category labels as clusters.
CLSClusteringP2P.v2	clustering	Clustering of titles + abstract from CLS dataset. Clustering of 13 sets on the main category.
StackOverflowQA	retrieval	The dataset is a collection of natural language queries and their corresponding response which may include some text mixed with code snippets. The task is to retrieve the most relevant response for a given query.
TwitterHjerneRetrieval	retrieval	Danish question asked on Twitter with the Hashtag #Twitterhjerne ('Twitter brain') and their corresponding answer.
AILAStatutes	retrieval	This dataset is structured for the task of identifying the most relevant statutes for a given situation.
ArguAna	retrieval	ArguAna: Retrieval of the Best Counterargument without Prior Topic Knowledge
HagridRetrieval	retrieval	HAGRID (Human-in-the-loop Attributable Generative Retrieval for Information-seeking Dataset)is a dataset for generative information-seeking scenarios. It consists of queriesalong with a set of manually labelled relevant passages
LegalBenchCorporateLobbying	retrieval	The dataset includes bill titles and bill summaries related to corporate lobbying.
LEMBPasskeyRetrieval	retrieval	passkey subset of dwzhu/LongEmbed dataset.
SCIDOCS	retrieval	SciDocs, a new evaluation benchmark consisting of seven document-level tasks ranging from citation prediction, to document classification and recommendation.
SpartQA	retrieval	Measuring the ability to retrieve the groundtruth answers to reasoning task queries on SpartQA.
TempReasonL1	retrieval	Measuring the ability to retrieve the groundtruth answers to reasoning task queries on TempReason l1.
TRECCOVID	retrieval	TRECCOVID is an ad-hoc search challenge based on the COVID-19 dataset containing scientific articles related to the COVID-19 pandemic.
WinoGrande	retrieval	Measuring the ability to retrieve the groundtruth answers to reasoning task queries on winogrande.
BelebeleRetrieval	retrieval	Belebele is a multiple-choice machine reading comprehension (MRC) dataset spanning 122 language variants (including 115 distinct languages and their scripts)
MLQARetrieval	retrieval	MLQA (MultiLingual Question Answering) is a benchmark dataset for evaluating cross-lingual question answering performance. MLQA consists of over 5K extractive QA instances (12K in English) in SQuAD format in seven languages - English, Arabic, German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA is highly parallel, with QA instances parallel between 4 different languages on average.
StatcanDialogueDatasetRetrieval	retrieval	A Dataset for Retrieving Data Tables through Conversations with Genuine Intents, available in English and French.
WikipediaRetrievalMultilingual	retrieval	The dataset is derived from Cohere's wikipedia-2023-11 dataset and contains synthetically generated queries.
CovidRetrieval	retrieval	COVID-19 news articles
Core17InstructionRetrieval	retrieval	Measuring retrieval instruction following ability on Core17 narratives for the FollowIR benchmark.
News21InstructionRetrieval	retrieval	Measuring retrieval instruction following ability on News21 narratives for the FollowIR benchmark.
Robust04InstructionRetrieval	retrieval	Measuring retrieval instruction following ability on Robust04 narratives for the FollowIR benchmark.
KorHateSpeechMLClassification	classification	The Korean Multi-label Hate Speech Dataset, K-MHaS, consists of 109,692 utterances from Korean online news comments, labelled with 8 fine-grained hate speech classes (labels: Politics, Origin, Physical, Age, Gender, Religion, Race, Profanity) or Not Hate Speech class. Each utterance provides from a single to four labels that can handles Korean language patterns effectively. For more details, please refer to the paper about K-MHaS, published at COLING 2022. This dataset is based on the Korean online news comments available on Kaggle and Github. The unlabeled raw data was collected between January 2018 and June 2020. The language producers are users who left the comments on the Korean online news platform between 2018 and 2020.
MalteseNewsClassification	classification	A multi-label topic classification dataset for Maltese News Articles. The data was collected from the press_mt subset from Korpus Malti v4.0. Article contents were cleaned to filter out JavaScript, CSS, & repeated non-Maltese sub-headings. The labels are based on the category field from this corpus.
MultiEURLEXMultilabelClassification	classification	EU laws in 23 EU languages containing annotated labels for 21 EUROVOC concepts.
BrazilianToxicTweetsClassification	classification	ToLD-Br is the biggest dataset for toxic tweets in Brazilian Portuguese, crowdsourced by 42 annotators selected from a pool of 129 volunteers. Annotators were selected aiming to create a plural group in terms of demographics (ethnicity, sexual orientation, age, gender). Each tweet was labeled by three annotators in 6 possible categories: LGBTQ+phobia, Xenophobia, Obscene, Insult, Misogyny and Racism.
CEDRClassification	classification	Classification of sentences by emotions, labeled into 5 categories (joy, sadness, surprise, fear, and anger).
CTKFactsNLI	pair-classification	Czech Natural Language Inference dataset of around 3K evidence-claim pairs labelled with SUPPORTS, REFUTES or NOT ENOUGH INFO veracity labels. Extracted from a round of fact-checking experiments.
SprintDuplicateQuestions	pair-classification	Duplicate questions from the Sprint community.
TwitterURLCorpus	pair-classification	Paraphrase-Pairs of Tweets.
ArmenianParaphrasePC	pair-classification	asparius/Armenian-Paraphrase-PC
indonli	pair-classification	IndoNLI is the first human-elicited Natural Language Inference (NLI) dataset for Indonesian. IndoNLI is annotated by both crowd workers and experts.
OpusparcusPC	pair-classification	Opusparcus is a paraphrase corpus for six European language: German, English, Finnish, French, Russian, and Swedish. The paraphrases consist of subtitles from movies and TV shows.
PawsXPairClassification	pair-classification	PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification
RTE3	pair-classification	Recognising Textual Entailment Challenge (RTE-3) aim to provide the NLP community with a benchmark to test progress in recognizing textual entailment
XNLI	pair-classification
PpcPC	pair-classification	Polish Paraphrase Corpus
TERRa	pair-classification	Textual Entailment Recognition for Russian. This task requires to recognize, given two text fragments, whether the meaning of one text is entailed (can be inferred) from the other text.
WebLINXCandidatesReranking	retrieval	WebLINX is a large-scale benchmark of 100K interactions across 2300 expert demonstrations of conversational web navigation. The reranking task focuses on finding relevant elements at every given step in the trajectory.
AlloprofReranking	retrieval	This dataset was provided by AlloProf, an organisation in Quebec, Canada offering resources and a help forum curated by a large number of teachers to students on all subjects taught from in primary and secondary school
VoyageMMarcoReranking	retrieval	a hard-negative augmented version of the Japanese MMARCO dataset as used in Voyage AI Evaluation Suite
WikipediaRerankingMultilingual	retrieval	The dataset is derived from Cohere's wikipedia-2023-11 dataset and contains synthetically generated queries.
RuBQReranking	retrieval	Paragraph reranking based on RuBQ 2.0. Give paragraphs that answer the question higher scores.
T2Reranking	retrieval	T2Ranking: A large-scale Chinese Benchmark for Passage Ranking
GermanSTSBenchmark	semantic-similarity	Semantic Textual Similarity Benchmark (STSbenchmark) dataset translated into German. Translations were originally done by T-Systems on site services GmbH.
SICK-R	semantic-similarity	Semantic Textual Similarity SICK-R dataset
STS12	semantic-similarity	SemEval-2012 Task 6.
STS13	semantic-similarity	SemEval STS 2013 dataset.
STS14	semantic-similarity	SemEval STS 2014 dataset. Currently only the English dataset
STS15	semantic-similarity	SemEval STS 2015 dataset
STSBenchmark	semantic-similarity	Semantic Textual Similarity Benchmark (STSbenchmark) dataset.
FaroeseSTS	semantic-similarity	Semantic Text Similarity (STS) corpus for Faroese.
FinParaSTS	semantic-similarity	Finnish paraphrase-based semantic similarity corpus
JSICK	semantic-similarity	JSICK is the Japanese NLI and STS dataset by manually translating the English dataset SICK (Marelli et al., 2014) into Japanese.
IndicCrosslingualSTS	semantic-similarity	This is a Semantic Textual Similarity testset between English and 12 high-resource Indic languages.
SemRel24STS	semantic-similarity	SemRel2024 is a collection of Semantic Textual Relatedness (STR) datasets for 14 languages, including African and Asian languages. The datasets are composed of sentence pairs, each assigned a relatedness score between 0 (completely) unrelated and 1 (maximally related) with a large range of expected relatedness values.
STS17	semantic-similarity	Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation
STS22.v2	semantic-similarity	SemEval 2022 Task 8: Multilingual News Article Similarity. Version 2 filters updated on STS22 by removing pairs where one of entries contain empty sentences.
STSES	semantic-similarity	Spanish test sets from SemEval-2014 (Agirre et al., 2014) and SemEval-2015 (Agirre et al., 2015)
STSB	semantic-similarity	A Chinese dataset for textual relatedness
MIRACLRetrievalHardNegatives	retrieval	MIRACL (Multilingual Information Retrieval Across a Continuum of Languages) is a multilingual retrieval dataset that focuses on search across 18 different languages. The hard negative version has been created by pooling the 250 top documents per query from BM25, e5-multilingual-large and e5-mistral-instruct.

How to evaluate on this benchmark

import mteb

benchmark = mteb.get_benchmark("MTEB(Multilingual, v2)")
model = mteb.get_model(YOUR_MODEL)
results = mteb.run(model, tasks=benchmark)

Citation

@article{enevoldsen2025mmtebmassivemultilingualtext,
  author = {Kenneth Enevoldsen and Isaac Chung and Imene Kerboua and Márton Kardos and Ashwin Mathur and David Stap and Jay Gala and Wissam Siblini and Dominik Krzemiński and Genta Indra Winata and Saba Sturua and Saiteja Utpala and Mathieu Ciancone and Marion Schaeffer and Gabriel Sequeira and Diganta Misra and Shreeya Dhakal and Jonathan Rystrøm and Roman Solomatin and Ömer Çağatan and Akash Kundu and Martin Bernstorff and Shitao Xiao and Akshita Sukhlecha and Bhavish Pahwa and Rafał Poświata and Kranthi Kiran GV and Shawon Ashraf and Daniel Auras and Björn Plüster and Jan Philipp Harries and Loïc Magne and Isabelle Mohr and Mariya Hendriksen and Dawei Zhu and Hippolyte Gisserot-Boukhlef and Tom Aarsen and Jan Kostkan and Konrad Wojtasik and Taemin Lee and Marek Šuppa and Crystina Zhang and Roberta Rocca and Mohammed Hamdy and Andrianos Michail and John Yang and Manuel Faysse and Aleksei Vatolin and Nandan Thakur and Manan Dey and Dipam Vasani and Pranjal Chitale and Simone Tedeschi and Nguyen Tai and Artem Snegirev and Michael Günther and Mengzhou Xia and Weijia Shi and Xing Han Lù and Jordan Clive and Gayatri Krishnakumar and Anna Maksimova and Silvan Wehrli and Maria Tikhonova and Henil Panchal and Aleksandr Abramov and Malte Ostendorff and Zheng Liu and Simon Clematide and Lester James Miranda and Alena Fenogenova and Guangyu Song and Ruqiya Bin Safi and Wen-Ding Li and Alessia Borghini and Federico Cassano and Hongjin Su and Jimmy Lin and Howard Yen and Lasse Hansen and Sara Hooker and Chenghao Xiao and Vaibhav Adlakha and Orion Weller and Siva Reddy and Niklas Muennighoff},
  doi = {10.48550/arXiv.2502.13595},
  journal = {arXiv preprint arXiv:2502.13595},
  publisher = {arXiv},
  title = {MMTEB: Massive Multilingual Text Embedding Benchmark},
  url = {https://arxiv.org/abs/2502.13595},
  year = {2025},
}

---

This benchmark card was automatically generated using MTEB