Institutional Repository of Vilnius University
Regos-kalbos modelių adaptavimo metodų tyrimas dirbant su lietuviškais duomenimis /
Abstract
dc:descriptionThis paper analyses the adaptation of vision-language models for the Lithuanian language. Since most vision-language models are trained and evaluated with English data, there is limited information about their ability to generate image descriptions in Lithuanian. This paper compares open-weight vision-language models in the task of generating image descriptions in Lithuanian. The models are also adapted, and quantization and pruning methods are applied to them. The obtained results show that vision-language models adapted using the LoRA method can improve the quality of image description generation in the Lithuanian language, quantization can reduce memory usage by 26% - 45%, depending on the selected vision-language model, but decrease token generation speed. Pruning can increase token generation speed by 53% - 519%, although its effect depends on the selected vision-language model and selected data set.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jarukaitis, Simonas,
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:ELABAETD308336576&prefLang=en_US