Datasets:
IPIS: Inclusive Polish Instruction Set
Note: The IPIS dataset is currently restricted, as it is being used for PolEval 2025 (Task 2). It will be publicly available here once the shared task ends.
IPIS is a collection of Polish instructions designed to improve the gender sensitivity and inclusiveness of large language models (LLMs). The IPIS instructions can be used to instruction-tune LLMs and are specifically designed to solve two tasks:
- gender-inclusive proofreading in Polish (IPIS-proofreading)
- Polish↔English gender-sensitive translation (see IPIS-translation)
IPIS data source
IPIS is built on a gender-inclusive text corpus manually annotated in the PLLuM project. The corpus comprises open-access texts written in standard Polish, a predominantly non-inclusive language, and representing diverse genres (see Data sources). All non-inclusive expressions in these texts are manually transformed into gender-sensitive versions (see Annotation procedure) by the IPI annotation team.
Dataset sources
European legal acts (Polish, English[):
- Charter of Fundamental Rights of the European Union
- The Treaty on Europen Union (Consolidated Version)
- The Treaty on the Functioning of the European Union (Consolidated Version)
Polish legal acts (Polish):
- Konstytucja Rzeczypospolitej Polskiej (Eng. Constitution of Poland)
Institutional documents (Polish):
- Regulamin pracy IPI PAN (Eng. Work regulation in ICS PAS)
- Status IPI PAN (Eng. ICS PAS Statute)
- Zasady podziału dotacji na działalność badawczo-rozwojową służącą rozwojowi młodych naukowców i uczestników studiów doktoranckich w IPI PAN (Eng. Rules for the distribution of subsidies for research and development activities aimed at the development of young scientists and PhD students at ICS PAS)
- Regulamin Rady Naukowej IPI PAN (Eng. Regulations of the Scientific Council of ICS PAS)
- Regulamin Szkoły Doktorskiej TIB PAN (Eng. Regulations of the Doctoral School TIB PAN)
- Tryb Przeprowadzania przed Radę Naukową IPI PAN Czynności w Postępowaniu w sprawie Nadania Stopnia Doktora (Eng. Procedure of Awarding the PhD Degree by the Scientific Council of ICS PAS)
- Regulamin zarządzania prawami autorskimi w IPI PAN (Eng. Copyright Managmenent Regulations at ICS PAS)
A balanced corpus of written texts (Polish)
- National Corpus of Polish (NLPre-based split of NKJP1M)
A subset of ParlaMint, i.e. a corpus of transcribed parliamentary debates (Polish)
- ParlaMint (year 2022)
Annotation procedure
Two annotators manually transform text units containing non-inclusive expressions into their gender-sensitive variants. Text passages that are already gender-neutral remain unchanged. When both annotated versions are valid (i.e. accepted by a super-annotator), they are included in the dataset to provide a more diverse range of training examples.
Possible gender-inclusive alternatives of the masculine generic form posłowie (Eng. deputies):
posłanki i posłowie
posłowie i posłanki
posłowie/posłanki
posłanki/posłowie
posł*owie/anki
Consequently, each text unit may have none (unchanged text), one, or two inclusive alternatives.
Annotation team
- dr Alina Wróblewska (leader),
- dr Anna Bulińska,
- Wiktora Gumienna,
- Martyna Lewandowska,
- Zuzanna Przybysz,
- Aleksandra Tomaszewska,
- Joanna Wołoszyn.
Credits
If you use IPIS in your research, please cite it as follows:
@misc{wróblewska2025integratinggenderinclusivitylarge,
title={Integrating gender inclusivity into large language models via instruction tuning},
author={Alina Wróblewska and Bartosz Żuk},
year={2025},
eprint={2508.18466},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2508.18466},
}
Acknowledgments
The IPIS dataset was created as part of the Universal Discourse: a multilingual model of discourse relations project, founded by the Polish National Science Centre (Grant number: 2023/50/A/HS2/00559).
Licence
IPIS is released under CC-BY-SA-4.0.
© Institute of Computer Science PAS, Alina Wróblewska
- Downloads last month
- 21