---

The Belebele Benchmark for Massively Multilingual NLU Evaluation

Belebele is a multiple-choice machine reading comprehension (MRC) dataset spanning 122 language variants. This dataset enables the evaluation of mono- and multi-lingual models in high-, medium-, and low-resource languages. Each question has four multiple-choice answers and is linked to a short passage from the FLORES-200 dataset. The human annotation procedure was carefully curated to create questions that discriminate between different levels of generalizable language comprehension and is reinforced by extensive quality checks. While all questions directly relate to the passage, the English dataset on its own proves difficult enough to challenge state-of-the-art language models. Being fully parallel, this dataset enables direct comparison of model performance across all languages. Belebele opens up new avenues for evaluating and analyzing the multilingual abilities of language models and NLP systems.

Please refer to our paper for more details, The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants.

Composition

• 900 questions per language variant
• 488 distinct passages, there are 1-2 associated questions for each.
• For each question, there is 4 multiple-choice answers, exactly 1 of which is correct.
• 122 language/language variants (including English).
• 900 x 122 = 109,800 total questions.

---

Download

Belebele can be downloaded here which you can download programmatically with the following bash command:

wget --trust-server-names https://dl.fbaipublicfiles.com/belebele/Belebele.zip

The dataset can additionally be used via the HuggingFace repo, such as with the datasets library.

Formatting details

• The link and split uniquely identifies a passage.
• The combination of passage (link and split) and question_number (either 1 or 2) uniquely identifies a question.
• The language of each row is denoted in the dialect column with the FLORES-200 code (see Languages below)
• The correct_answer_num is one-indexed (e.g. a value of 2 means mc_answer2 is correct)

Multimodal Belebele

In the 2M-belebele directory, you will find details about the multimodal (speech+ASL) version of belebele.

---

Pausible Evaluation Settings

Thanks to the parallel nature of the dataset and the simplicity of the task, there are many possible settings in which we can evaluate language models. In all evaluation settings, the metric of interest is simple accuracy (# correct / total). Several of the evaluation settings were implemented with several models and discussed in the paper.

Evaluating models on Belebele in English can be done via finetuning, few-shot, or zero-shot. For other target languages, we propose the incomprehensive list of evaluation settings below. Settings that are compatible with evaluating non-English models (monolingual or cross-lingual) are denoted with ^.

No finetuning

• Zero-shot with natural language instructions (English instructions)
  • For chat-finetuned models, we give it English instructions for the task and the sample in the target language in the same input.
  • See the sample_zero_shot_instructions.md file for specific details on how we evaluated instructed models for the paper.
• Zero-shot with natural language instructions (translated instructions)^
  • Same as above, except the instructions are translated to the target language so that the instructions and samples are in the same language. The instructions can be human or machine-translated.
• Few-shot in-context learning (English examples)
  • A few samples (e.g. 5) are taken from the English training set (see below) and prompted to the model. Then, the model is evaluated with the same template but with the passages, questions, and answers in the target language.
  • For our experiments, we use the template: P: <passage> \n Q: <question> \n A: <mc answer 1> \n B: <mc answer 2> \n C: <mc answer 3> \n D: <mc answer 4> \n Answer: <Correct answer letter>. We perform prediction by picking the answer within [A, B, C, D] that has the highest probability relatively to the others.
• Few-shot in-context learning (translated examples)^
  • Same as above, except the samples from the training set are translated to the target language so that the examples and evaluation data are in the same language. The training samples can be human or machine-translated.

With finetuning

• English finetune & multilingual evaluation
  • The model is finetuned to the task using the English training set, probably with a sequence classification head. Then the model is evaluated in all the target languages individually.
• English finetune & cross-lingual evaluation
  • Same as above, except the model is evaluated in a cross-lingual setting, where for each question, the passage & answers could be provided in a different language. For example, passage could be in language x, question in language y, and answers in language z.
• Translate-train^
  • For each target language, the model is individually finetuned on training samples that have been machine-translated from English to that language. Each model is then evaluated in the respective target language.
• Translate-train-all
  • Similar to above, except here the model is trained on translated samples from all target languages at once. The single finetuned model is then evaluated on all target languages.
• Translate-train-all & cross-lingual evaluation
  • Same as above, except the single finetuned model is evaluated in a cross-lingual setting, where for each question, the passage & answers could be provided in a different language.
• Translate-test
  • The model is finetuned using the English training data and then the evaluation dataset is machine-translated to English and evaluated on the English.
  • This setting is primarily a reflection of the quality of the machine translation system, but is useful for comparison to multilingual models.

In addition, there are 83 additional languages in FLORES-200 for which questions were not translated for Belebele. Since the passages exist in those target languages, machine-translating the questions & answers may enable decent evaluation of machine reading comprehension in those languages.

---

Training Set

As discussed in the paper, we also provide an assembled training set consisting of samples

The Belebele dataset is intended to be used only as a test set, and not for training or validation. Therefore, for models that require additional task-specific training, we instead propose using an assembled training set consisting of samples from pre-existing multiple-choice QA datasets in English. We considered diverse datasets, and determine the most compatible to be RACE, SciQ, MultiRC, MCTest, MCScript2.0, and ReClor.

For each of the six datasets, we unpack and restructure the passages and questions from their respective formats. We then filter out less suitable samples (e.g. questions with multiple correct answers). In the end, the dataset comprises 67.5k training samples and 3.7k development samples, more than half of which are from RACE. We provide a script (assemble_training_set.py) to reconstruct this dataset for anyone to perform task finetuning.

Since the training set is a joint sample of other datasets, it is governed by a different license. We do not claim any of that work or datasets to be our own. See the Licenses section.

---

Languages in Belebele

FLORES-200 Code	English Name	Script	Family
acm_Arab	Mesopotamian Arabic	Arab	Afro-Asiatic
afr_Latn	Afrikaans	Latn	Germanic
als_Latn	Tosk Albanian	Latn	Paleo-Balkanic
amh_Ethi	Amharic	Ethi	Afro-Asiatic
apc_Arab	North Levantine Arabic	Arab	Afro-Asiatic
arb_Arab	Modern Standard Arabic	Arab	Afro-Asiatic
arb_Latn	Modern Standard Arabic (Romanized)	Latn	Afro-Asiatic
ars_Arab	Najdi Arabic	Arab	Afro-Asiatic
ary_arab	Moroccan Arabic	Arab	Afro-Asiatic
arz_Arab	Egyptian Arabic	Arab	Afro-Asiatic
asm_Beng	Assamese	Beng	Indo-Aryan
azj_Latn	North Azerbaijani	Latn	Turkic
bam_Latn	Bambara	Latn	Mande
ben_Beng	Bengali	Beng	Indo-Aryan
ben_Latn^	Bengali (Romanized)	Latn	Indo-Aryan
bod_Tibt	Standard Tibetan	Tibt	Sino-Tibetan
bul_Cyrl	Bulgarian	Cyrl	Balto-Slavic
cat_Latn	Catalan	Latn	Romance
ceb_Latn	Cebuano	Latn	Austronesian
ces_Latn	Czech	Latn	Balto-Slavic
ckb_Arab	Central Kurdish	Arab	Iranian
dan_Latn	Danish	Latn	Germanic
deu_Latn	German	Latn	Germanic
ell_Grek	Greek	Grek	Hellenic
eng_Latn	English	Latn	Germanic
est_Latn	Estonian	Latn	Uralic
eus_Latn	Basque	Latn	Basque
fin_Latn	Finnish	Latn	Uralic
fra_Latn	French	Latn	Romance
fuv_Latn	Nigerian Fulfulde	Latn	Atlantic-Congo
gaz_Latn	West Central Oromo	Latn	Afro-Asiatic
grn_Latn	Guarani	Latn	Tupian
guj_Gujr	Gujarati	Gujr	Indo-Aryan
hat_Latn	Haitian Creole	Latn	Atlantic-Congo
hau_Latn	Hausa	Latn	Afro-Asiatic
heb_Hebr	Hebrew	Hebr	Afro-Asiatic
hin_Deva	Hindi	Deva	Indo-Aryan
hin_Latn^	Hindi (Romanized)	Latn	Indo-Aryan
hrv_Latn	Croatian	Latn	Balto-Slavic
hun_Latn	Hungarian	Latn	Uralic
hye_Armn	Armenian	Armn	Armenian
ibo_Latn	Igbo	Latn	Atlantic-Congo
ilo_Latn	Ilocano	Latn	Austronesian
ind_Latn	Indonesian	Latn	Austronesian
isl_Latn	Icelandic	Latn	Germanic
ita_Latn	Italian	Latn	Romance
jav_Latn	Javanese	Latn	Austronesian
jpn_Jpan	Japanese	Jpan	Japonic
kac_Latn	Jingpho	Latn	Sino-Tibetan
kan_Knda	Kannada	Knda	Dravidian
kat_Geor	Georgian	Geor	kartvelian
kaz_Cyrl	Kazakh	Cyrl	Turkic
kea_Latn	Kabuverdianu	Latn	Portuguese Creole
khk_Cyrl	Halh Mongolian	Cyrl	Mongolic
khm_Khmr	Khmer	Khmr	Austroasiatic
kin_Latn	Kinyarwanda	Latn	Atlantic-Congo
kir_Cyrl	Kyrgyz	Cyrl	Turkic
kor_Hang	Korean	Hang	Koreanic
lao_Laoo	Lao	Laoo	Kra-Dai
lin_Latn	Lingala	Latn	Atlantic-Congo
lit_Latn	Lithuanian	Latn	Balto-Slavic
lug_Latn	Ganda	Latn	Atlantic-Congo
luo_Latn	Luo	Latn	Nilo-Saharan
lvs_Latn	Standard Latvian	Latn	Balto-Slavic
mal_Mlym	Malayalam	Mlym	Dravidian
mar_Deva	Marathi	Deva	Indo-Aryan
mkd_Cyrl	Macedonian	Cyrl	Balto-Slavic
mlt_Latn	Maltese	Latn	Afro-Asiatic
mri_Latn	Maori	Latn	Austronesian
mya_Mymr	Burmese	Mymr	Sino-Tibetan
nld_Latn	Dutch	Latn	Germanic
nob_Latn	Norwegian Bokmål	Latn	Germanic
npi_Deva	Nepali	Deva	Indo-Aryan
npi_Latn^	Nepali (Romanized)	Latn	Indo-Aryan
nso_Latn	Northern Sotho	Latn	Atlantic-Congo
nya_Latn	Nyanja	Latn	Afro-Asiatic
ory_Orya	Odia	Orya	Indo-Aryan
pan_Guru	Eastern Panjabi	Guru	Indo-Aryan
pbt_Arab	Southern Pashto	Arab	Indo-Aryan
pes_Arab	Western Persian	Arab	Iranian
plt_Latn	Plateau Malagasy	Latn	Austronesian
pol_Latn	Polish	Latn	Balto-Slavic
por_Latn	Portuguese	Latn	Romance
ron_Latn	Romanian	Latn	Romance
rus_Cyrl	Russian	Cyrl	Balto-Slavic
shn_Mymr	Shan	Mymr	Kra-Dai
sin_Latn^	Sinhala (Romanized)	Latn	Indo-Aryan
sin_Sinh	Sinhala	Sinh	Indo-Aryan
slk_Latn	Slovak	Latn	Balto-Slavic
slv_Latn	Slovenian	Latn	Balto-Slavic
sna_Latn	Shona	Latn	Atlantic-Congo
snd_Arab	Sindhi	Arab	Indo-Aryan
som_Latn	Somali	Latn	Afro-Asiatic
sot_Latn	Southern Sotho	Latn	Atlantic-Congo
spa_Latn	Spanish	Latn	Romance
srp_Cyrl	Serbian	Cyrl	Balto-Slavic
ssw_Latn	Swati	Latn	Atlantic-Congo
sun_Latn	Sundanese	Latn	Austronesian
swe_Latn	Swedish	Latn	Germanic
swh_Latn	Swahili	Latn	Atlantic-Congo
tam_Taml	Tamil	Taml	Dravidian
tel_Telu	Telugu	Telu	Dravidian
tgk_Cyrl	Tajik	Cyrl	Iranian
tgl_Latn	Tagalog	Latn	Austronesian
tha_Thai	Thai	Thai	Kra-Dai
tir_Ethi	Tigrinya	Ethi	Afro-Asiatic
tsn_Latn	Tswana	Latn	Atlantic-Congo
tso_Latn	Tsonga	Latn	Afro-Asiatic
tur_Latn	Turkish	Latn	Turkic
ukr_Cyrl	Ukrainian	Cyrl	Balto-Slavic
urd_Arab	Urdu	Arab	Indo-Aryan
urd_Latn^	Urdu (Romanized)	Latn	Indo-Aryan
uzn_Latn	Northern Uzbek	Latn	Turkic
vie_Latn	Vietnamese	Latn	Austroasiatic
war_Latn	Waray	Latn	Austronesian
wol_Latn	Wolof	Latn	Atlantic-Congo
xho_Latn	Xhosa	Latn	Atlantic-Congo
yor_Latn	Yoruba	Latn	Atlantic-Congo
zho_Hans	Chinese (Simplified)	Hans	Sino-Tibetan
zho_Hant	Chinese (Traditional)	Hant	Sino-Tibetan
zsm_Latn	Standard Malay	Latn	Austronesian
zul_Latn	Zulu	Latn	Atlantic-Congo

^ denotes a language variant not in FLORES-200

Further Stats

• 122 language variants, but 115 distinct languages (ignoring scripts)
• 27 language families
• 29 scripts
• Avg. words per passage = 79.1 (std = 26.2)
• Avg. sentences per passage = 4.1 (std = 1.4)
• Avg. words per question = 12.9(std = 4.0)
• Avg. words per answer = 4.2 (std = 2.9)

---

License

The Belebele dataset is licensed under the license found in the LICENSE_CC-BY-SA4.0 file in the root directory of this source tree.

The training set and assembly code is, however, licensed differently. The majority of the training set (data and code) is licensed under CC-BY-NC, however portions of the project are available under separate license terms: NLTK is licensed under the Apache 2.0 license; pandas and NumPy are licensed under the BSD 3-Clause License.

Citation

If you use this data in your work, please cite:

@inproceedings{bandarkar-etal-2024-belebele,
    title = "The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants",
    author = "Bandarkar, Lucas  and
      Liang, Davis  and
      Muller, Benjamin  and
      Artetxe, Mikel  and
      Shukla, Satya Narayan  and
      Husa, Donald  and
      Goyal, Naman  and
      Krishnan, Abhinandan  and
      Zettlemoyer, Luke  and
      Khabsa, Madian",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand and virtual meeting",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.44",
    pages = "749--775",
}