Projects of OPI PIB

HIVE AI: Development and pilot implementation of large language models in Polish public administration

The goal of the project is to support the digitalisation and automation of routine administrative tasks in Poland through the pilot implementation of Polish large language models (LLMs) in public administration systems.

Project data

Project start date2025-02-20
Project end date2025-12-31
Project cost18983055 PLN

Lider

  • Research and Academic Computer Network

Partners

  • AGH University of Krakow
  • Centre for Information Technology
  • Institute of Computer Science of the Polish Academy of Sciences
  • Institute of Slavic Studies of the Polish Academy of Sciences
  • National Information Processing Institute
  • Wrocław University of Science and Technology
  • University of Lodz

HIVE AI is a direct continuation of the Polish Large Language Model (PLLuM) project. HIVE AI aims to facilitate the integration of Polish language models into the mObywatel app, develop a clerical assistant for the Polish Ministry of Digital Affairs, and conduct a pilot implementation of LLMs in a municipal or provincial office.
To implement LLMs and adapt them to specific needs of public administration identified during the pilot phase, new linguistic datasets will be collected. Additionally, new instruction fine-tuning datasets will be developed, comprising prompts (queries or commands) and expected responses. Depending on the type of instruction, datasets may include additional elements such as reasoning, context, and keywords. New datasets for aligning LLMs with human preferences will also be created, containing sets of prompts and multiple response variations that are assessed for quality.

Based on these resources, the PLLuM model family will be expanded to comprise new models of various sizes, including:

  • General-purpose models, which are designed for a wide range of linguistic tasks, with a particular focus on administrative environment
  • Generators, which are specialised Retrieval-Augmented Generation (RAG) models whose objective is to generate final responses based on natural language queries and retrieved contextual documents that are most relevant to particular queries.

The project will involve close collaboration with public institutions in which the pilot implementation will roll out, ensuring efficient support and automation of their operations.

Funding

The project is funded by the Ministry of Digital Affairs as part of budget item 27 ‘Informatisation in 2025’.