We release the first representative subset of the PLLuM Instruction Corpus (PLLuMIC), which we believe to be useful in guiding and planning the development of similar LLM datasets. PLLuMIC, at its core, is a hand-crafted set of LLM fine-tuning Polish language instructions. The corpus is described in more detail in a forthcoming paper titled The PLLuM Instruction Corpus. We plan regular updates and significant extensions of the corpus.

https://arxiv.org/abs/2511.17161

@misc{pęzik2025plluminstructioncorpus,
      title={The PLLuM Instruction Corpus}, 
      author={Piotr Pęzik and Filip Żarnecki and Konrad Kaczyński and Anna Cichosz and Zuzanna Deckert and Monika Garnys and Izabela Grabarczyk and Wojciech Janowski and Sylwia Karasińska and Aleksandra Kujawiak and Piotr Misztela and Maria Szymańska and Karolina Walkusz and Igor Siek and Maciej Chrabąszcz and Anna Kołos and Agnieszka Karlińska and Karolina Seweryn and Aleksandra Krasnodębska and Paula Betscher and Zofia Cieślińska and Katarzyna Kowol and Artur Wilczek and Maciej Trzciński and Katarzyna Dziewulska and Roman Roszko and Tomasz Bernaś and Jurgita Vaičenonienė and Danuta Roszko and Paweł Levchuk and Paweł Kowalski and Irena Prawdzic-Jankowska and Marek Kozłowski and Sławomir Dadas and Rafał Poświata and Alina Wróblewska and Katarzyna Krasnowska-Kieraś and Maciej Ogrodniczuk and Michał Rudolf and Piotr Rybak and Karolina Saputa and Joanna Wołoszyn and Marcin Oleksy and Bartłomiej Koptyra and Teddy Ferdinan and Stanisław Woźniak and Maciej Piasecki and Paweł Walkowiak and Konrad Wojtasik and Arkadiusz Janz and Przemysław Kazienko and Julia Moska and Jan Kocoń},
      year={2025},
      eprint={2511.17161},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.17161}, 
}

The PLLuMIC dataset is available here: https://huggingface.co/datasets/pelcra/PLLuMIC The data is divided into two subsets: main organic part and synthetic extension.

The organic samples were carefully curated by human annotators, developed in line with the annotation guidelines and covering a functional typology. The synthetic extension was created using a strong, permissively licensed LLM (DeepSeek v3) and a custom pipeline incorporating organic samples injection.

Total instructions: 1,278

All instructions were annotated by professional annotators. Each sample was developed in accordance with comprehensive annotation guidelines and subsequently reviewed by a senior annotator to ensure full compliance with quality standards. The annotation process followed a functional typology designed to encompass key areas of model competence.

Total instructions: 54,921

Each type and subtype has been handled individually, with careful attention to quality standards and guidelines. Each synthetic sample was generated by injecting suitable organic examples, with differentiation measures applied to ensure diversity. There are currently no system prompts in the subset, but there is an ongoing work to include them in the nearest future.

All subtypes within these types are covered. The thematic categorisation is yet to come in future updates.

The PLLuMIC dataset is distributed as a JSONL file storing rows with conversations between a user and an AI assistant. There are 2 JSONL files included, one for the organic component and one for the synthetic extension. Each conversation is a JSON structure described by following fields:

• dataset_name: Name of the dataset (PLLuMIC).
• dataset_source: Source organization (CLARIN-BIZ-bis).
• conv_id: Unique identifier for the conversation (3242183cbce2).
• messages: Array of dialogue messages (user/assistant/system exchanges).

Each entry in messages contains:

• instruction_id: Unique ID for the instruction/task (2a07c2eca0cb).
• seq: Sequence number (-1 for system, 0,1,2,… for user/assistant turns).
• role: Speaker role (system, user, or assistant).
• content: Text of the message (empty for some system prompts).
• type: Interaction type (e.g., Dialog, Generation).
• subtype: List of task subtype (e.g., [System prompt, Text simplification]).
• topic: List of relevant topics (e.g., [Geography]).
• language: Language code (e.g., pol for Polish).
• source: References (e.g., Wikipedia URLs).