Belebele
MetaBelebele is a multiple-choice reading comprehension benchmark in 122 language variants, with 900 questions per variant written about short FLORES-200 passages. Every question has four answers, one correct, and the set is fully parallel so scores compare directly across languages. Annotators wrote the questions in English under quality checks before translation. Meta FAIR released it.
Openness
5 high confidence- license
- cc-by-sa-4.0 (Covers the benchmark; the assembled English training set is mostly CC-BY-NC per the README.)
- access
- public (Ungated on Hugging Face and as a zip from Meta's file server.)
- dataset_card
- present
The benchmark carries a share-alike license and downloads without a gate. The optional English training set assembled from other datasets is mostly non-commercial, but that part is not the benchmark.
- https://huggingface.co/api/datasets/facebook/belebele recorded 2026-09-24
gated: false; license tag cc-by-sa-4.0.
- https://raw.githubusercontent.com/facebookresearch/belebele/main/README.md recorded 2026-09-24
"The Belebele dataset is licensed under the license found in the LICENSE_CC-BY-SA4.0 file"; "The majority of the training set (data and code) is licensed under CC-BY-NC".
Adoption
3 high confidenceHugging Face downloads over the trailing month for the Belebele repository. Downloads of the zip from Meta's file server are not counted.
- https://huggingface.co/api/datasets/facebook/belebele recorded 2026-09-24
84563 downloads in the trailing 30 days for facebook/belebele
Capability
3 medium confidenceIn the typical language Belebele targets, its deepest test has four hundred to sixteen hundred items that count toward depth. An item count says how much a set can test, not how precisely it ranks models.
- https://arxiv.org/abs/2207.04672 recorded 2026-11-08
"we evaluated the performance of over 40,000 different translation directions using a human-translated benchmark, Flores-200"
- https://arxiv.org/html/2308.16884 recorded 2026-11-08
"For all languages included in the corpus, the context passages were taken directly from the FLoRes -200 dataset"; "We opted to create multiple-choice questions and answers in English and then translate"; "Belebele was created end-to-end without the use of machine translation technology, relying solely on experts fluent in English and the target language."; for the romanized variants, "We thus had annotators transliterate from the native to Latin script"
- https://datasets-server.huggingface.co/size?dataset=facebook/belebele recorded 2026-11-08
the Hub's split sizes for facebook/belebele: one test split of 900 rows in each of the 122 language-variant configs, 109,800 in all
- https://raw.githubusercontent.com/EleutherAI/lm-evaluation-harness/main/lm_eval/tasks/README.md recorded 2026-11-08
Task table row: "belebele ... Language understanding tasks in a variety of languages and scripts ... Multiple (122 languages)".
- https://raw.githubusercontent.com/facebookresearch/belebele/main/README.md recorded 2026-11-08
"900 questions per language variant"; "122 language/language variants"; "900 x 122 = 109,800 total questions".
Verified 2026-09-24