Exploring N-Shot Learning Approaches in Generative Large Language Models for the Serbian Language

Fine-tuning large language models for a specific task requires a lot of resources and time. Also, to successfully fine-tune the model, we need to have a high-quality dataset. The lack of finetuning datasets is particularly present in low-resource languages, such as Serbian. Therefore, there is a need to examine other approaches for adapting a large language model to a specific task. In this paper, we investigated the N-shot learning capabilities of generative models for sentiment analysis of reviews in the Serbian language. We used two models for N -shot learning - the non-opensource GPT-4o and the open-source Mistral model. We fine-tuned the BERTić model and tested it on the same test set for comparison purposes. The GPT-40 model achieved excellent results, slightly better than the BERTić model. It turned out that the Mistral model lags behind the GPT-4o model and the standard BERTić finetuning.