pinzhenchen/Cultivar-flores

🪴 Cultivar

Cultivar is a locale-oriented evaluation benchmark for multilingual machine translation. The name "Cultivar-flores" highlights that the benchmark is a cultivated variety of FLORES through localisation.

Motivation

Imagine you test translation systems or LLMs on Spanish-to-English translation, this input is not your true use case: "La facultad de medicina de la Universidad de Stanford ..." (the Stanford University School of Medicine ...). Furthermore, since the input has been translated from English, which was originally taken from Wikimedia projects, your model might have seen the answer.

Our benchmark, Cultivar lets you have: "La Facultad de Medicina de la Universidad Complutense de Madrid ..." (the Complutense University of Madrid School of Medicine ...), which is comparable, localised, and uncontaminated. Why not check out the samples to get a feel?

Dataset Details

Dataset Description

Cultivar is a multilinguallocale translation benchmark released under CC BY-SA 4.0. For each locale, the dataset has 200 sentences in the dev split, following the FLORES naming. As of now (v1.0), the benchmark covers 28 language-location combinations (locales).

Cultivar has been created by localising subsets of FLORES to different locations. Please see our paper for more technical details.

Intended Use

Cultivar is intended as a translation benchmark with localised test instances. You should use the dev split for testing. You can use Cultivar alone to get locale-oriented evaluation results, i.e. your evaluation unit will be language pairs for a certain locale, for instance, cmn_Hans_China->Eng_Latn.

Since it has been created by localising FLORES instances, Cultivar can be paired with FLORES at the instance level (and language-locale level) to provide two diagnostics: FLORES contamination and localisation robustness, independently of translation capability.

Python code

As agreed, please 1) do not redistribute the data publicly unless you protect it from automatic scraping; 2) do not optimise your system towards this benchmark.

Note: Pass a subset name when loading the data, because otherwise the default is a 1-instance-per-locale sample subset.

from datasets import get_dataset_config_names, load_dataset

dataset_name = "pinzhenchen/Cultivar-flores"

# Load a specific subset,
#   which is formatted as: {lang1}_{Script1}_{Location}-{lang2}_{Script2}
dataset = load_dataset(
    dataset_name,
    name="hin_Deva_India-eng_Latn",
)

# Load all subsets
all_subsets = {
    subset: load_dataset(dataset_name, name=subset)
    for subset in get_dataset_config_names(dataset_name)
    if subset != "sample"
}
dev_data = all_subsets["hin_Deva_India-eng_Latn"]["dev"]

Dataset Structure

Taking the deu_Latn_Switzerland-eng_Latn subset as an example, each instance has the content as follows:

{"id": "183",
 "flores_id": "768",
 "lang1_iso_639_3_code": "deu",
 "lang1_iso_15924_code": "Latn",
 "lang1_glottocode": "stan1295",
 "lang2_iso_639_3_code": "eng",
 "lang2_iso_15924_code": "Latn",
 "lang2_glottocode": "stan1293",
 "location": "Switzerland",
 "lang1_text": "Weitere Skiveranstaltungen finden im Skigebiet Corviglia in St. Moritz statt, etwa 200 km von Bern entfernt.",
 "lang2_text": "Other skiing events will be at the Corviglia ski area in St. Moritz, about 200 km from Bern.",
 "last_updated": "1.0",
 "split": "dev"
}

As a comparison, the corresponding entries in the original FLORES for deu_Latn and eng_Latn are:

{
  "deu": "Weitere Skiveranstaltungen finden im Skigebiet Taizicheng in Zhangjiakou statt, etwa 220 km (140 Meilen) von Peking entfernt.",
  "eng": "Other skiing events will be at the Taizicheng ski area in Zhangjiakou, about 220 km (140 miles) from Beijing."
}

Data Fields

• id: ID number for each instance of data. Instances with the same ID in the same split are comparable instances of each other.
• flores_id: the ID number of the corresponding (original, un-localised) instance in FLORES. You can retrieve more metadata (e.g. url, domain, etc) for this instance via the original FLORES.
• lang1_iso_639_3: The ISO 639-3 code of language 1 (the localised language).
• lang1_iso_15924: The ISO 15924 indicating the writing script for language 1.
• lang1_glottocode: The Glottocode corresponding to language 1.
• lang2_iso_639_3: The ISO 639-3 code of language 2 (the paired localised language).
• lang2_iso_15924: The ISO 15924 indicating the writing script for language 2.
• lang2_glottocode: The Glottocode corresponding to language 2.
• location: The location this instance has been localised to. All instances in the same subset have the same location, although there can be rare exceptions.
• lang1_text: The localised text in language 1.
• lang2_text: The paired localised text in language 2.
• last_updated: The Cultivar-flores version where the given instance was last updated.
• split: The data split, same as the one for the corresponding instance in the original FLORES.

Questions or contributing?

Fixes and new localisation contributions are always welcome. We are working on setting this up.

Here is a list of technical and data contributors (sorted by contribution then by first name):

• v1.0: Pinzhen Chen, Koel Dutta Chowdhury, Xiaoya Xu, David Tan, Doreen Osmelak, Ona de Gibert, Ariun-Erdene Tumurchuluun, Ashok Urlana, Bouazza Laracha, Fedor Sizov, Hale Sirin, Jesujoba Alabi, Karrar Talib Abed, Mateusz Klimaszewski, Nikolay Bogoychev, Niyati Bafna, Patricia Schmidtova, Preksha Manjunath Shanbhag, Sherrie Shen, Teresa Sy Ortin, Vilem Zouhar, Vivek Iyer, Yasser Hamidullah, Yusser Al Ghussin, and Zheng Zhao.

Please feel free to get in touch using the email address provided in the associated paper.

Citation

Please cite our work as:

@article{chen-etal-2026-cultivar,
      title={Cultivar: A Contrastive and Locale-Oriented Translation Benchmark for Investigating Contamination and Localisation Robustness}, 
      author={Pinzhen Chen and Koel Dutta Chowdhury and Xiaoya Xu and David Tan and Doreen Osmelak and Ona de Gibert and Ariun-Erdene Tumurchuluun and Ashok Urlana and Fedor Sizov and Hale Sirin and Jesujoba Alabi and Karrar Talib Abed and Mateusz Klimaszewski and Nikolay Bogoychev and Niyati Bafna and Patricia Schmidtova and Preksha Manjunath Shanbhag and Sherrie Shen and Vilem Zouhar and Vivek Iyer and Yasser Hamidullah and Yusser Al Ghussin and Zheng Zhao},
      year={2026},
      journal={arXiv preprint},
      url={https://arxiv.org/abs/2608.09766}, 
}

Please also cite FLORES.