Adapting large language models to low-resource languages presents three interconnected challenges: inefficient tokenization, scarcity of high-quality annotated data, and limited resources for instruction tuning. We present a reproducible approach that addresses each challenge using data-centric methods that primarily rely on unlabeled text corpora, parallel translation data, and a multilingual base model. Our approach combines (1) vocabulary surgery for tokenizer adaptation without full retraining, (2) cross-lingual transfer of quality classifiers via translation, enabling filtering without target-language annotations, and (3) generation of instruction data through translation, task conversion, and targeted synthesis. We validate this recipe by adapting Gemma-3-12B to Ukrainian. %, producing Lapa-12BOur pretrained model achieves top performance on Ukrainian benchmarks, while our instruction-tuned variant demonstrates strong performance on translation (33 BLEU on FLORES), summarization, and question-answering tasks, while requiring 1.5x fewer tokens than the original model for the same text. We release all models, datasets, classifiers, and code to enable replication for other languages.

Text preprocessing is a fundamental component of high-quality speech synthesis. This work presents a novel rule-based phonemizer combined with a sentence-level lexical stress prediction model to improve phonetic accuracy and prosody prediction in the text-to-speech pipelines. We also introduce a new benchmark dataset with annotated stress patterns designed for evaluating lexical stress prediction systems at the sentence level.Experimental results demonstrate that the proposed phonemizer achieves a 1.23% word error rate on a manually constructed pronunciation dataset, while the lexical stress prediction pipeline shows results close to dictionary-based methods, outperforming existing neural network solutions.