FairForget/StereoSet-UK-Eval

StereoSet-UK Eval

StereoSet-UK Eval contains 949 Ukrainian masked triplets derived from the intrasentence portion of the StereoSet development set. The Ukrainian full sentences were translated with the DeepL API and received technical cleanup. English source text is omitted.

The shared Ukrainian templates and fills were cut mechanically from the translated sentences. All three candidates reconstruct their corresponding full sentence exactly. Each retained template passed the automated frame-alignment rules: its fixed-frame ratio is within 0.20 of the English source ratio, the number of mask occurrences is unchanged, and the Ukrainian fills do not share context word forms or stems absent from the English fills.

The 79 source targets use unique, standardized Ukrainian labels. Aggregation by target_uk therefore preserves the original target boundaries used for LMS, SS, and ICAT.

This is a partial evaluation subset. Report scores as results on the 949-item StereoSet-UK Eval subset. Full StereoSet development-set scores require complete development-set coverage.

Fields

Field	Description
item_id	Original StereoSet item identifier
bias_type_uk	Ukrainian bias-category label
target_uk	Canonical Ukrainian label for the bias target
template_uk	Shared Ukrainian sentence containing [BLANK]
stereotype_fill_uk	Ukrainian stereotype completion
anti_stereotype_fill_uk	Ukrainian anti-stereotype completion
unrelated_fill_uk	Ukrainian unrelated completion

For evaluation, score all three fills in the same template_uk. This is a provisional machine-translated research dataset. Human validation is still required.

Ukrainian adaptation

The Ukrainian translation, cleaning, and adaptation were prepared by Viktoriia Makovska, Oleksandra Ostafiichuk, and Tetiana Zakharchenko.

Source and license

This dataset is derived from StereoSet at source commit ead7d086a64a192a1eca88e0dd2fd163de375218. The source and this translation use the Creative Commons Attribution-ShareAlike 4.0 International license.

Citation

Please cite both the Ukrainian adaptation and the original StereoSet paper.

Ukrainian adaptation

@dataset{makovska_etal_2026_stereoset_uk_eval,
  title = {{StereoSet-UK Eval}: Ukrainian Translation, Cleaning, and Masked Evaluation Adaptation},
  author = {Makovska, Viktoriia and Ostafiichuk, Oleksandra and Zakharchenko, Tetiana},
  year = {2026},
  publisher = {FairForget},
  url = {https://huggingface.co/datasets/FairForget/StereoSet-UK-Eval}
}

Original StereoSet

@inproceedings{nadeem-etal-2021-stereoset,
  title = {{StereoSet}: Measuring stereotypical bias in pretrained language models},
  author = {Nadeem, Moin and Bethke, Anna and Reddy, Siva},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing},
  year = {2021},
  pages = {5356--5371},
  doi = {10.18653/v1/2021.acl-long.416},
  url = {https://aclanthology.org/2021.acl-long.416/}
}