Institutional Repository of Vilnius University
Regos-kalbos modelių adaptavimas medicininiams aprašams lietuvių kalba /
Abstract
dc:descriptionThis thesis investigates whether existing vision–language models can be further finetuned to improve their performance on medical data in the Lithuanian language. The relevance of this work is determined bythefact that mostmodernartificialintelligence models are primarily oriented toward the English language, which limits their applicability to less commonly used languages such as Lithu anian, especially in the medical domain. For this reason, there is a growing need to developmethods that enable the effective adaptationofexistingmodelsforprocessing and generatingLithuanian me dical data.During the research, experiments were conducted using the MedGemma and Gemma 4 vision–language models together with the PathVQA, VQARAD, and SLAKE datasets. The aim of the experiments was to improve these opensource models for generating medical responses in Lithu anian and to evaluate their ability to understand medical images and provide accurate information. The Unsloth framework was used for additional finetuning of the models, enabling more efficient training while reducing the amount of computational resources required.In addition, an evaluation methodologywasdevelopedtoprovideamorecomprehensiveanalysisofthemodels’performance. The evaluation considered not only overall statistical accuracy, but also the correct usage of medical terminology and the grammatical correctness of the Lithuanian language. In this way, the study ai medtoassessnotonlythetechnical performance of the models, but also their practical applicability in real medical scenarios.The results of the research demonstrated that both models can be success fully finetuned and that their generated responses regarding medical data in Lithuanian improve after training. However, it was observed that the MedGemma model achieved better results both before and after additional finetuning, such as an 82.2% accuracy on yes/no question answering for the finetuned SLAKE dataset. This model generated medical terminology more accurately and used grammatically more correct Lithuanian language. The obtained results indicate that specialized me dical vision–language models have strong potential for application in the Lithuanian medical domain, while additional finetuning can significantly improve their quality and practical usability.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kisieliūtė, Rūta,
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:ELABAETD308337479&prefLang=en_US