Lapa LLM
Data-Efficient Adaptation of Multilingual LLMs to Ukrainian
How did we build Lapa LLM for Ukrainian?
The question
Why this work
This paper presents the data pipeline behind Lapa LLM, an openly released Ukrainian language model built on Gemma-3-12B. It tackles inefficient tokenization, limited high-quality data, and scarce instruction examples. The project gives the Ukrainian NLP community reusable models, training data, quality classifiers, and code.
Lapa LLM is named in honor of Ukrainian researcher Valentyn Lapa. Project background ↗
The approach
A visual guide
Three changes to the data pipeline
Vocabulary surgery
Improve Ukrainian tokenization without retraining the whole model from scratch.
Quality filtering
Transfer quality signals across languages through translated data.
Instruction data
Combine translation, task conversion, and targeted synthesis.
What the work contributes
The adapted model uses about two-thirds as many tokens for the same text as the original model—the paper reports 1.5× fewer tokens. It also reports strong Ukrainian benchmark and instruction-task results, and releases models, data, classifiers, and code.
Original paper & resources
In Proceedings of the Fifth Ukrainian Natural Language Processing Conference (UNLP 2026), pages 155-168, Lviv, Ukraine. Association for Computational Linguistics.
Citation
Paniv, Y., Didenko, B., Haltiuk, M., Humennyy, V., Kravchenko, A., Kyslyi, R., Makovska, V., Orlovskyi, A., Ruban, B., Rudko, M.-Y., Senyk, A., Drushchak, N., Chaplynskyi, D., & Romanyshyn, M. (2026). Data-Efficient Adaptation of Multilingual LLMs to Ukrainian. In Proceedings of the Fifth Ukrainian Natural Language Processing Conference (UNLP 2026), 155-168. Association for Computational Linguistics. https://aclanthology.org/2026.unlp-1.14/
BibTeX
@inproceedings{paniv-etal-2026-data,
title = {Data-Efficient Adaptation of Multilingual {LLM}s to {U}krainian},
author = {Paniv, Yurii and Didenko, Bohdan and Haltiuk, Mykola and Humennyy, Vladyslav and Kravchenko, Andrian and Kyslyi, Roman and Makovska, Viktoriia and Orlovskyi, Artem and Ruban, Bohdan and Rudko, Maksym-Yurii and Senyk, Anastasiia and Drushchak, Nazarii and Chaplynskyi, Dmytro and Romanyshyn, Mariana},
booktitle = {Proceedings of the Fifth {U}krainian Natural Language Processing Conference ({UNLP} 2026)},
year = {2026},
address = {Lviv, Ukraine},
publisher = {Association for Computational Linguistics},
pages = {155--168},
url = {https://aclanthology.org/2026.unlp-1.14/}
}