PLLuM, a set of Polish artificial intelligence models, will be piloted this year in the Ministry of Digital Affairs and a selected government office, according to Dr. Agnieszka Karlińska, who will lead the HIVE consortium. She noted that one of the biggest challenges was training these AI models with expert input.
This week, the Ministry of Digital Affairs introduced PLLuM—18 AI models designed for businesses, public administration, and researchers—along with a chatbot that is freely available to all users.
“HIVE is a follow-up to the PLLuM project, focused on implementation and running until the end of 2025. We are expanding the consortium by adding two additional institutions: the Central IT Center (COI) and Cyfronet AGH,” said Dr. Karlińska from the AI Security Research Center at NASK. She explained that COI will provide implementation support, while Cyfronet, which operates Poland’s largest supercomputer, will contribute computing power.
AI Assistants for Government and Citizens
The PLLuM models will be developed for use in government offices, specifically in the form of “administrative assistants” and an assistant for the mObywatel mobile application. These AI-powered assistants will function as chatbots specialized in handling inquiries from and to government officials, as well as answering citizens’ questions within mObywatel, Karlińska explained.
The Ministry of Digital Affairs plans to launch the first pilot implementation of the models within its own structures this year. “We also intend to test them in a selected government office before making them available in additional institutions,” she added.
She emphasized that the availability of lightweight AI models allows for deployment even in smaller government offices. This is made possible by utilizing the Mixture of Experts (MoE) architecture, which optimizes the number of active parameters, improving efficiency while reducing resource consumption.
Source: Adobe Stock
Challenges in Developing Poland’s AI Models
One of the main challenges in developing these Polish AI models was collecting a sufficiently large dataset in the Polish language. “We ensured that data was acquired ethically and legally. We began with a thorough legal analysis, particularly focusing on Polish and EU regulations concerning copyright protection. A major emphasis was placed on obtaining data directly from publishers through licensing agreements. For publicly available sources, we verified the licenses and any restrictions related to text and data mining (TDM),” Karlińska stated.
Another challenge was fine-tuning and “training” the models using specific instructions and preference-based learning. These were designed to be “organic,” meaning they were developed manually by experts.
Instructions consisted of query-response sets (known as prompts), which helped the AI model learn how to perform specific tasks, such as composing emails. “Manually crafting these instructions involved linguists and AI specialists writing prompts as users would, then designing ideal responses and complete dialogues,” Karlińska explained.
Preferences, on the other hand, involved training sets that included a user query, a preferred (better) response, and a rejected (worse) response. “With preference-based learning, we aimed to teach models to generate answers that were both accurate and helpful while maintaining a balanced and safe tone. Evaluations considered not only factual accuracy but also neutrality and general safety,” she highlighted.
The research team assessed whether model-generated responses were free of potentially harmful content, including offensive language, socially inappropriate statements, or legally and ethically questionable information.
Karlińska emphasized that the manual fine-tuning of Poland’s AI models was a time-consuming process, taking several months and requiring extensive coordination. “There are no clear guidelines on how to create AI training instructions and preferences, nor any established typology for them.”
“Developers of foreign AI models typically do not disclose details on this process. We had to take an experimental approach. I believe we have achieved good results,” she concluded.
The PLLuM and HIVE Projects
The Polish AI models were developed within the PLLuM consortium, which included Wrocław University of Science and Technology (the project’s initial leader), the Institute of Computer Science of the Polish Academy of Sciences, the Institute of Slavic Studies of the Polish Academy of Sciences, the National Research Institute NASK, the National Information Processing Institute (OPI-PIB), and the University of Łódź.
As the PLLuM consortium’s work was originally scheduled to conclude at the end of 2024, the Ministry of Digital Affairs announced the formation of the new HIVE consortium, which will further develop Polish AI and integrate it into public administration. Dr. Agnieszka Karlińska from the AI Security Research Center at NASK will lead this initiative.
Data collection for PLLuM began in February 2024. Model training took place between June and September, with additional fine-tuning in October and optimization in November. Initially, the model was scheduled for public release in December.
The project is funded by the Ministry of Digital Affairs, which owns the results and oversees its development. In 2024, the ministry allocated 14.5 million PLN to the project, with 19 million PLN secured for 2025.
Monika Blandyna Lewkowicz (PAP)
Read full article (PL): Ekspert: w tym roku modele polskiej sztucznej inteligencji zostaną pilotażowo wdrożone w MC | Nauka w Polsce
