The authors examine the concept of datasets intended for machine learning of algorithms / synthetic neuronet models, i. e. artificial intelligence (AI) models. Russian Federation pursues the Strategy for Information Society for 2017–2030, with the goal to build knowledge society. Today the libraries have been accumulating, preserving and disseminating knowledge, and in the context of datasets, providing the strategy concept. The authors define dataset as a structured data array organized to solve particular problems: the dataset structure, in its turn, depends on the primary data type, character and formality. In the article, the authors substantiate this statement and discuss two basic technological models for building datasets (based on text corpus).