The vast amount of observations needed to train new generation AI models (Foundation Models) necessitates a strategy of combining data from multiple repositories in a semi-automatic way to minimize human involvement. However, many public data sources present challenges such as inhomogeneity, lack of machine-actionable data, and manual access barriers. These issues can be mitigated through the consequent adherence to the FAIR (Findable, Accessible, Interoperable, Reusable) data principles, as well as state-of-the-art data standards and tools. In the poster, we highlight the inhomogeneity of the schema definitions in the field, provide helpful tips on what could improve the AI-readiness of data and inspect example data sources which implement the most novel concepts in working with data and metadata in the machine-actionable fashion.