Institutional Repository of Vilnius University
Didelių kalbos modelių lietuvių kalbai tyrimas /
Abstract
dc:descriptionIn recent years, large language models have become one of the fastest-developing areas of technology, yet their capabilities for low-resource languages such as Lithuanian remain insufficient. This work investigates the effectiveness of LoRA adaptation for Lithuanian morphological analysis and accentuation tasks, using the Ministral 3 3B model as the base model. Training data was collected and prepared from three sources: LRT, Wikisources, and the General Corpus of the Lithuanian Language (BLKT). Four LoRA adapters were trained, differing in data volume, diversity, and LoRA rank. Results show that specialized LoRA adaptation significantly improves model performance in accentuation, whose accuracy increased from 8.0% to 80.0%, and overall morphological analysis, whose accuracy rose from 33.7% to 56.5%. Furthermore, the adapted three billion parameter model was able to substantially outperform GPT5.5 and Gemini 3.1 models in both tasks. Error analysis revealed that the model’s accentuation accuracy reaches approximately 96% when it produces an output, with most errors stemming from word omission rather than incorrect accentuation. The study also found that data diversity has a greater impact on adaptation quality than data quantity, and that improper LoRA parameter selection, specifically an excessively high rank combined with a high learning rate, can cause overfitting even with large datasets.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sazonik, Jekaterina,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308338032&prefLang=en_US