Back to Gap Map Model components / Language-specific evaluation sets

FLORES+

Open Language Data Initiative
gated / Overall score: 3.0

FLORES+ is a multilingual machine translation benchmark of sentences sampled from Wikinews, Wikijunior and Wikivoyage and translated from English into more than 200 language varieties, with dev and devtest splits aligned across every language. It continues Meta's FLORES-200, and a separate blind test set stays with Meta. The Open Language Data Initiative maintains it and accepts community fixes and new languages.

Openness

3 high confidence
3.0
license
cc-by-sa-4.0
access
auto (Automatic Hugging Face gate: agree not to re-host where crawlers can reach it and to keep its contents out of training data when evaluating.)
dataset_card
present

The sentences carry a share-alike license and the gate approves anyone who accepts its terms. Those terms ask users not to re-host the files where crawlers can find them and to keep them out of training data, to protect the benchmark from contamination.

Adoption

3 high confidence
3.0

Hugging Face downloads over the trailing month for the one FLORES+ repository. Copies bundled inside evaluation harnesses and earlier FLORES-200 downloads from Meta are not counted.

Capability

3 medium confidence
3.0

In the typical language FLORES+ targets, its deepest test has four hundred to sixteen hundred items that count toward depth. Of its 215 target languages, 188 have a countable test. An item count says how much a set can test, not how precisely it ranks models.

  • https://aclanthology.org/2022.tacl-1.30.pdf recorded 2026-11-08

    "The dev set is meant to be used for hyperparameter tuning. The devtest is meant to be used for testing purposes during the development phase. The test set will not be released"

  • https://arxiv.org/abs/2205.12446 recorded 2026-11-08

    FLEURS "is an n-way parallel speech dataset in 102 languages built on top of the machine translation FLoRes-101 benchmark".

  • 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://huggingface.co/datasets/openlanguagedata/flores_plus recorded 2026-11-08

    "For each language, the dataset has 997 sentences for the dev split and 1012 sentences for the devtest split"; "The data consists of translations primarily from English into over 200 language varieties"; "The separate blind test set, originally developed by Meta, is not managed by OLDI and not part of this repository"; "Currently 230 language varieties"; the coverage table marks Dargwa, Udmurt and Southern Uzbek "dev only"

  • https://raw.githubusercontent.com/facebookresearch/belebele/main/README.md recorded 2026-11-08

    "Each question has four multiple-choice answers and is linked to a short passage from the FLORES-200 dataset."

  • https://raw.githubusercontent.com/facebookresearch/flores/main/flores200/README.md recorded 2026-11-08

    "FLORES-200 has several languages which were not translated from English. Specifically, several languages were translated from Spanish, French, Russian and Modern Standard Arabic."; the README's language table lists each FLORES-200 code, ace_Arab to zul_Latn, and its FLORES-101 map pairs "arb_Arab | ara"

Verified 2026-09-24