Vladislav Poritski

We use BelarusianGLUE, a recently introduced benchmark, to analyze the performance of open-weight large language models (LLMs) on Belarusian language understanding tasks. The impact of prompting language, few-shot prompts, orthography (modern/classical/Latin), chat templates, and evaluation mode (discriminative/ generative) is investigated. Our findings suggest that more recent models generally perform better, but improvements are gradual. Fine-tuning on related Slavic languages doesn’t always improve Belarusian understanding. Classical orthography has limited impact, while latinization degrades performance. Analysis of specific tasks (sentiment analysis, Winograd schema challenge) reveals biases in the models, difficulties with understanding linguistic structure, and gaps in world knowledge and cultural context.

Subject: e-Scripta

Keywords: natural language processing Belarusian language large language models language understanding evaluation