HIVE AI – Development and pilot deployment of large language models in Polish public administration

Project factsheet

Project summary

The goal of HIVE AI is to support the digitalisation and automation of routine administrative tasks in Poland through the pilot deployment of Polish large language models in public administration systems. The project is a continuation of PLLuM – Polish Large Language Model: building on the experience gained there, it moves from a strictly research phase to a deployment phase. Its main operational goal is to adapt language models to the needs of public administration.

Project activities

The project consists of four complementary activities:

1. Language data corpora: collecting new text data, especially data that is not publicly available and documents from the administrative domain, and preparing instruction datasets for fine-tuning and preference data for alignment, based on the project’s own typology.
2. Model training: extending the PLLuM family with models of several sizes – general-purpose models adapted to various natural language processing tasks with a strong focus on administration, as well as generative models, including retrieval-augmented generation (RAG) models. Training covers pretraining, fine-tuning and alignment.
3. Security and evaluation: developing tools to assess the quality and security of language models in government use cases, including input/output filtering algorithms and output correction mechanisms that reduce harmful or unwanted content.
4. Pilot deployment: preparing the deployment of Polish language models in the mObywatel app (including an mObywatel chatbot) and in selected public institutions, such as the Ministry of Digital Affairs, through virtual assistant pilots.

Consortium

The project is carried out by a consortium of eight research centres and institutions implementing IT and cybersecurity projects for the public sector, coordinated by NASK. The team brings together computer scientists, linguists, lawyers, sociologists and cybersecurity experts.

• NASK – National Research Institute (consortium leader)
• AGH University of Krakow
• Centralny Ośrodek Informatyki (Central Information Technology Centre)
• Institute of Computer Science, Polish Academy of Sciences (represented on the consortium board by Maciej Ogrodniczuk)
• Institute of Slavic Studies, Polish Academy of Sciences
• National Information Processing Institute – National Research Institute (OPI PIB)
• Wrocław University of Science and Technology
• University of Łódź

Results

• PLLuM-12B-nc-250715 – the first publicly available result of the HIVE AI consortium, announced in July 2025. It updates the PLLuM-12B-nc model released in February 2025 and was trained on a better-curated dataset extended with government texts, including the Public Information Bulletin (BIP). It is available in base, instruct and chat variants.
• The PLLuM models developed by the consortium are available on Hugging Face and can be tried out in the PLLuM chat; a prompt book explains how to use the chat effectively. More information is available on the PLLuM website.
• In October 2025, the City of Gdynia and the HIVE AI consortium signed an agreement on the pilot deployment of PLLuM models in the city’s chatbot and in its Public Information Bulletin search engine.
• On 31 December 2025, a virtual assistant based on PLLuM was made available to users of the mObywatel app.

See also: PLLuM at Science with the Palace in the Background.