A PhD student at the Applied Sciences Faculty of Ukrainian Catholic University, Yurii Paniv, is leading a project to develop an open large language model for the Ukrainian language — Lapa LLM v0.1.2. It is one of the most efficient large language models (LLMs) optimized for processing Ukrainian-language content.
A large language model (LLM) is a computer program that has learned to work with language as if it “understands” it. Its task is to predict the next word in a sentence, formulate thoughts correctly, or answer questions. To learn this, a model like this “reads” millions or even billions of texts — from books and articles to social media posts. It does not memorize everything literally, but instead learns patterns: how people usually speak, how they build sentences, and which words tend to be used together.
The model is called “large” because it contains an enormous number of internal parameters — “settings” that help it predict how to formulate correct answers. Thanks to this complex system, it can build logical sentences, respond to questions, and even stylistically imitate human speech. The more parameters it has, the more accurate, flexible, and naturally it works. Such models are the basis of chatbots (for example, ChatGPT), automatic translators, voice assistants, and many other applications.
At the end of October, a team of Ukrainian researchers from the Applied Sciences Faculty of Ukrainian Catholic University, AGH University of Krakow, Igor Sikorsky Kyiv Polytechnic Institute, and Lviv Polytechnic presented Lapa LLM — an advanced open large language model based on Gemma-3-12B with a focus on processing the Ukrainian language. The team worked on a volunteer basis, in their free time.
The team included: Yurii Paniv, Andrian Kravchenko, Bohdan Didenko, Mykola Haltiuk, Vladyslav Humennyi, Maksym-Yurii Rudko, Bohdan Ruban, Artem Orlovskyi, Anastasiia Senyk, Mariana Romanyshyn, Dmytro Chaplynskyi, Roman Kyslyi, Nazarii Drushchak, and Nikita Trynus.
“We started with UCU servers that had two A6000 graphics cards. We received these servers thanks to the support of Eleks through a memorial grant in honor of Oleksii Skrypnyk, and they are intended for shared use by PhD students. We prepared a detailed project plan, identified what needed to be done first (collecting data and creating tests to evaluate the model), and already presented this plan in various environments. Later, through acquaintances, we came across the French startup Comand AI. They were interested in a Ukrainian model capable of working with documents in a closed environment (mainly for military use cases). This was mutually beneficial: for them for their needs, and for us so that the model would be open and commercially available. Hugging Face supported us by providing a corporate subscription. This allowed us to easily work with and store datasets and models,” Yurii Paniv told dev.ua.
Lapa LLM is named after Valentyn Lapa — one of the creators of the Method of Group Accounting of Arguments, which became a precursor of modern deep learning. This is a way of teaching computers to independently find patterns in data, similar to how the human brain does.
The goal of the project is to create the best open Ukrainian-language artificial intelligence model and to publish high-quality data on which it will be trained.
In an interview with dev.ua, Yurii Paniv noted that the team achieved high tokenization efficiency — that is, how the model splits text into small parts it can work with. Thanks to a tokenizer adaptation method developed by Mykola Haltiuk, they managed to replace 80,000 tokens with Ukrainian ones without loss of quality. The result: Lapa LLM requires 1.5 times fewer tokens to work with Ukrainian, which ensures three times fewer computations and higher performance.
“To simplify: every large language model does not operate on text itself, but on predefined pieces of words (tokens), turning them into numbers. The better the text is compressed into these numbers, the faster the model will produce results and the better it can be trained, because the longer the context, the worse the model learns as a result. Changing the tokenizer itself is not a problem, but most known methods led to a loss of model quality,” Yurii Paniv explained.
The team’s second achievement is open datasets for different tasks.
“Thanks to high-quality open translation datasets, Lapa is the best translator in the English–Ukrainian pair. We conducted measurements across different domains, including encyclopedic text, social media, and work documents. We fine-tuned the model on legal translations using parallel corpora of EU and Ukrainian legislation. According to some estimates, if the adaptation of Ukrainian legislation to EU standards were done only by humans, it would take about nine years. Our model can significantly speed up this process by working in a secure, closed environment and producing excellent legal translations. We also fine-tuned the model for text summarization and question-answering from text. This is important for chatbots, document analysis, and RAG systems,” the researcher added.
According to test results, Lapa LLM shows performance comparable to MamayLM — currently the strongest Ukrainian-language model — and demonstrates very good results in text summarization, question answering, translation, and document work.
The language model is protected from Russian propaganda and disinformation.
“This was at the top of our priority list, and we approached the problem in several stages. First, we created several datasets based on sources where, for example, disinformation specialists had already annotated existing hostile narratives. One of the sources we used was the very well-structured VoxCheck, which contained pairs of ‘propaganda label’ and ‘counter-argument.’ We used these pairs to generate similar pairs using other models. That is, we had a narrative and two responses: the correct one based on facts and one from the propaganda point of view,” Yurii Paniv explained. “Next, we developed a classifier that gives a score of how true or false a given text is. Then we ran this tool through our large text corpus and removed all materials that contained propaganda. We used these pairs as conditionally disputable questions and correct, fact-based answers. Our tests confirmed that the model works well with disinformation, and we are satisfied with this result.”
You can try the model here:
huggingface.co/spaces/lapa-llm/lapa
