The Polish Large Language Model, PLLUm, trained on Polish texts, is being developed by the a consortium led by Wrocław University of Science and Technology. The project aims to enhance the innovation and competitiveness of the Polish scientific and technological sector. Scientists from the Department of Artificial Intelligence at WIT are working on the project.

PLLuM, which stands for Polish Large Language Universal Model, is a project involving a large Polish language model based on artificial intelligence technology. Its main goal is to revolutionise interactions with the Polish language in the digital world by providing more precise and contextually embedded natural language processing.

"The concept of the model was created with the idea of intensive use of data in the Polish language, allowing for a deeper understanding of the nuances and specifics of our language, as well as providing high-quality knowledge about our history and culture," says Dr. Eng. Jan Kocoń from the Department of Artificial Intelligence at WIT.

PLLuM, often compared to the famous ChatGPT, is a response to the growing demand for specialised language tools that can better serve Polish users in various applications, from education to customer service. The scientist adds that developing such a model will be a significant step toward more personalized and effective AI solutions that better meet the needs of the Polish-speaking society.

PLLuM thinks and speaks in Polish

Although ChatGPT manages quite well with the Polish language, "quite well" is often not sufficient, according to scientists from Wrocław University of Science and Technology, the National Research Institute NASK, the National Information Processing Institute, the Institute of Computer Science PAS, the University of Lodz, and the Institute of Slavic Studies PAS. Therefore, they decided to create a tool that, as one of the Polish Romantic era poets, Juliusz Słowacki, might say: "I want the flexible language to say everything the mind thinks..."

"The difference between PLLUm and ChatGPT mainly arises from their language specifications. PLLUm will be optimised to work with the Polish language, including better understanding of local expressions and idioms, as well as the cultural and historical specificities of Poland," explains Dr. Eng. Jan Kocoń. "ChatGPT, being a more universal model, lacks this depth of understanding of Polish specificity. This is crucial to creating natural and fluid conversations."

As a result, PLLUm can offer more precise and tailored responses in Polish, particularly important in areas requiring specialised knowledge or local context.

PLLuM open and free

However, the linguistic nuances are not the only focus. A scientist from the Department of Artificial Intelligence at WIT highlights another important aspect. "Even if ChatGPT already offers a service at a fairly good level of knowledge of the Polish language, we still don't have full control over it," notes Dr. Eng. Jan Kocoń. He emphasises that the intensive use of the famous American model is expensive. "And applying it to texts containing sensitive data is often excluded because we cannot process such data locally."

PLLuM will be an open tool, crucial given the growing demand for digital services in the Polish language. "The project aims to increase innovation and competitiveness in the Polish scientific and technological sector," emphasises the researcher from WIT. "Its development will contribute to a better understanding of the cultural and linguistic realities by machines, opening up new possibilities in education, public administration, and industry. Additionally, PLLUm aims to facilitate access to advanced language technologies for Polish users, regardless of their technical knowledge.

Due to its open source nature, PLLUm can be modified and customised for specific user needs. This can be particularly significant for Polish businesses and institutions looking for solutions adapted to the local language and culture, which can be deployed on their own servers. Collaboration with academic and research institutions can also contribute to the continuous development and improvement of the model. Dr. Eng. Jan Kocoń explains that PLLUm, being available to a wide range of users, has the potential to contribute significantly to the development of the Polish digital space.

PLLuM truthfulness

The PLLUm training process will rely on advanced machine learning techniques, utilising large datasets in the Polish language to ensure the high quality and relevance of generated responses, as well as credibility.

"During the model training process, diverse sources of texts will be used, from literature to internet articles, aiming to provide a comprehensive knowledge base. Special attention is given to avoiding the introduction of biased or false information into the model," emphasises Dr. Eng. Jan Kocoń. "The team working on PLLUm can introduce manual corrections in the form of instructions to ensure even better solution quality. The planned continuous model updating process is also crucial in a dynamically changing information environment."

According to the scientist, one of the biggest challenges in creating PLLUm is ensuring high-quality data for its training, which must be representative in a broad Polish context. Another challenge is maintaining a balance between the effectiveness and its hardware requirements and operational costs. Developing effective methods to evaluate and testing the model is also crucial to ensure its reliability and appropriate response to user queries, especially if the model generates incorrect responses.

Managing user expectations and educating them about the capabilities and limitations of the model is also essential. Finally, the challenge lies in continuously updating the model to keep up with evolving language and culture, requiring ongoing research and development.

PLLuM safe or risky?

Chat, capable of answering any question and generating any type of text, brings significant benefits and numerous possibilities, but models like Chat GPT or PLLUm also raise certain concerns.

"The threats associated with PLLUm and similar artificial intelligence models primarily concern privacy and data security issues. There is a risk that large datasets used for model training may contain sensitive information, raising concerns about their protection," says the scientist from the Department of Artificial Intelligence. "Additionally, there is a potential risk of using the model to generate disinformation, especially in political or social contexts. Another threat is the potential misuse of AI technology for unethical purposes, such as creating fake identities."

Dr. Eng. Jan Kocoń is aware that the automation of content may lead to a decline in the credibility of information on the internet and make it difficult to distinguish between content generated by humans and that created by machines. Increased reliance on automatic systems may also result in the loss of language skills and critical thinking in society. Finally, the development of such technologies may contribute to increased unemployment in certain sectors where human tasks can be replaced by automation.

„That is why it is crucial for the development and implementation of AI models, such as PLLUm, to be conducted with consideration of these risks," concludes one of the creators of PLLUm.