From LLM Generation to Knowledge Representation: Creating and Structuring the GeminiKnowledge-sr QA Dataset for Serbian

<p>LLMs capable of answering questions, fulfilling diverse user requests, and functioning as chatbots rely heavily on extensive datasets. However, for the Serbian language, there is a significant lack of high-quality datasets structured in a question-and-answer (QA) format. To address this gap, we have created the largest QA dataset GeminiKnowledge–sr in Serbian, foundational dataset specifically designed to improve QA capabilities in Serbian LLMs, which can work in conjunction with or as an alternative to formal KGs. This dataset is designed to encompass a broad spectrum of knowledge domains, ranging from specialized expert fields to areas concerning sports, culture, personal development, everyday activities, and more. The development started with generation of tree-level taxonomy: Area (26) → Topic (317) → Facet (6376), followed by generation of context, questions and answers. After filtering 127,091 QA pairs remained. In the next phase, the knowledge base question generation involved creation of natural language questions based on structured data, represented as fact triples.</p>