README.md

---
license: apache-2.0
task_categories:
- translation
language:
- rm
- de
configs:
  - config_name: de_DE-rm-rumgr
    data_files:
      - split: test
        path: de_DE-rm-rumgr.jsonl

  - config_name: de_DE-rm-sursilv
    data_files:
      - split: test
        path: de_DE-rm-sursilv.jsonl

  - config_name: de_DE-rm-sutsilv
    data_files:
      - split: test
        path: de_DE-rm-sutsilv.jsonl

  - config_name: de_DE-rm-surmiran
    data_files:
      - split: test
        path: de_DE-rm-surmiran.jsonl

  - config_name: de_DE-rm-puter
    data_files:
      - split: test
        path: de_DE-rm-puter.jsonl

  - config_name: de_DE-rm-vallader
    data_files:
      - split: test
        path: de_DE-rm-vallader.jsonl


---

# WMT24++ Reference Translations for Romansh

## Description  
WMT24++ benchmark in Romansh (six varieties: Rumantsch Grischun, Sursilvan, Sutsilvan, Surmiran, Puter, and Vallader).

- Paper: ["Expanding the WMT24++ Benchmark with Rumantsch Grischun, Sursilvan, Sutsilvan, Surmiran, Puter, and Vallader"](https://arxiv.org/abs/2509.03148)
- Code: https://github.com/ZurichNLP/romansh_mt_eval
- Original WMT24++ benchmark (55 languages): https://huggingface.co/datasets/google/wmt24pp

- The reference translation have been created by human translators based on the German version of WMT24++.

> [!NOTE] The Sutsilvan subset is currently incomplete. See Appendix A of the [paper](https://arxiv.org/abs/2509.03148) for an overview. We will update the dataset once the Sutsilvan translations have been completed.

## License  
<!-- Contributions to existing datasets must be released under the same license as the parent dataset. For completely new contributions, we encourage the use of an open license. At a minimum, data should be made available for research use. Please specify the license using an SPDX license identifier. -->  
Apache-2.0

## Language codes  
<!-- * If this language is assigned an ISO 639-3 individual language code (not a macrolanguage code), specify it here. * Please specify the script this language is written in using an ISO 15924 code. * If this language is assigned a Glottocode, please specify it here. -->

| Subset                         | Variety               | ISO 639-3 | ISO 15924 script code | Glottocode   |
|--------------------------------|-----------------------|-----------|-----------------------|--------------|
| **de_DE-rm-rumgr.jsonl**       | Rumantsch Grischun    | roh       | Latn                  | ruma1247     |
| **de_DE-rm-sursilv.jsonl**     | Sursilvan             | roh       | Latn                  | surs1244     |
| **de_DE-rm-sutsilv.jsonl**     | Sutsilvan             | roh       | Latn                  | suts1235     |
| **de_DE-rm-surmiran.jsonl**    | Surmiran              | roh       | Latn                  | surm1243     |
| **de_DE-rm-puter.jsonl**       | Puter                 | roh       | Latn                  | uppe1396     |
| **de_DE-rm-vallader.jsonl**    | Vallader              | roh       | Latn                  | lowe1386     |
  
## Data Creation Workflow  
<!-- What workflow was followed in creating this dataset? E.g., for a translated dataset, relevant information includes: what language the content was translated from, the number of translators, aggregate translator information (how many were native speakers in the target language, how many were highly proficient in the target languages, how many had professional translation experience), was any fraction of the data checked independently by third parties, etc. -->
The data acquisition process was structured into three steps to ensure high quality, consistency, and adherence to idiom-specific conventions.

1. **Translation**: We hired language professionals who are native speakers of both German and the respective Romansh idiom.
2. **Review**: Two expert linguists of Lia Rumantscha reviewed a sample of translations for a representative selection of varieties, and formulated feedback that was communicated to all translators.
3. **Revision**: The translators incorporated the feedback into the reference translations.

We provided the translators with a guidelines document, inspired by the WMT24 translator brief.

The key points of the guidelines, which we provide in Appendix E of the [paper](https://arxiv.org/abs/2509.03148), are:

* The German text is the main source for the translation into Romansh, while the English text can be used as an additional reference in case of ambiguity.
* No AI tools should be used for the translation.

The translators and reviewers had access to the complete context of each segment, including a link to the original website from which the segment was extracted (e.g., for segments from the Speech domain, the original YouTube video).

## Citations

This dataset:
```bibtex
@inproceedings{vamvas-et-al-2025-expanding,
  title        = {Expanding the {WMT}24++ Benchmark with {R}umantsch {G}rischun, {S}ursilvan, {S}utsilvan, {S}urmiran, {P}uter, and {V}allader},
  author       = {Vamvas, Jannis and P\'{e}rez Prat, Ignacio and Soliva, Not Battesta and Baltermia-Guetg, Sandra and Beeli, Andrina and Beeli, Simona and Capeder, Madlaina and Decurtins, Laura and Gregori, Gian Peder and Hobi, Flavia and Holderegger, Gabriela and Lazzarini, Arina and Lazzarini, Viviana and Rosselli, Walter and Vital, Bettina and Rutkiewicz, Anna and Sennrich, Rico},
  year={2025},
  eprint={2509.03148},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2509.03148},
}
```

WMT24:
```bibtex
@inproceedings{kocmi-etal-2024-findings,
    title = "Findings of the {WMT}24 General Machine Translation Shared Task: The {LLM} Era Is Here but {MT} Is Not Solved Yet",
    author = "Kocmi, Tom  and  Avramidis, Eleftherios  and  Bawden, Rachel  and  Bojar, Ond{\v{r}}ej  and  Dvorkovich, Anton  and  Federmann, Christian  and  Fishel, Mark  and  Freitag, Markus  and  Gowda, Thamme  and  Grundkiewicz, Roman  and  Haddow, Barry  and  Karpinska, Marzena  and  Koehn, Philipp  and  Marie, Benjamin  and  Monz, Christof  and  Murray, Kenton  and  Nagata, Masaaki  and  Popel, Martin  and  Popovi{\'c}, Maja  and  Shmatova, Mariya  and  Steingr{\'i}msson, Steinth{\'o}r  and  Zouhar, Vil{\'e}m",
    editor = "Haddow, Barry  and  Kocmi, Tom  and  Koehn, Philipp  and  Monz, Christof",
    booktitle = "Proceedings of the Ninth Conference on Machine Translation",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.wmt-1.1/",
    doi = "10.18653/v1/2024.wmt-1.1",
    pages = "1--46",
}
```

WMT24++:
```bibtex
@inproceedings{deutsch-etal-2025-wmt24,
    title = "{WMT}24++: Expanding the Language Coverage of {WMT}24 to 55 Languages {\&} Dialects",
    author = "Deutsch, Daniel  and  Briakou, Eleftheria  and  Caswell, Isaac Rayburn  and  Finkelstein, Mara  and  Galor, Rebecca  and  Juraska, Juraj  and  Kovacs, Geza  and  Lui, Alison  and  Rei, Ricardo  and  Riesa, Jason  and  Rijhwani, Shruti  and  Riley, Parker  and  Salesky, Elizabeth  and  Trabelsi, Firas  and  Winkler, Stephanie  and  Zhang, Biao  and  Freitag, Markus",
    editor = "Che, Wanxiang  and  Nabende, Joyce  and  Shutova, Ekaterina  and  Pilehvar, Mohammad Taher",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-acl.634/",
    doi = "10.18653/v1/2025.findings-acl.634",
    pages = "12257--12284",
    ISBN = "979-8-89176-256-5",
}
```