This paper proposes a methodology for database generation for Machine Learning-based power system stability studies. Power system operable region is a high-multidimensional space, therefore, creating a good quality database for training accurate predictive ML models is a challenging task. A sampling strategy aimed to explore high-multidimensional spaces avoiding curse-of-dimensionality is proposed. The strategy is based on sampling into the whole operating region ,detecting the region of the space with high information content and biasing the sampling toward that space. The computing of the entropy function is used for the stability margin detection and information content quantification. Latin Hypercube Sampling is adopted for sampling in the multidimensional space.The study is presented through the application to a small-signal stability problem. An essential power system, representing a grid with low synchronous generation rated power and high penetration of converter interfaced generation technologies, is used as case study to validate the methodology. Data are generated with the proposed methodology and with a conventional brute force approach. The two databases are compared and used for ML models training. Both the possibility to train regression and classification models is tested.