At present, a machine learning application is accomplished by carefully crafting a single table from an often complex, multi-table database. The metadata necessary to create this table is rarely formally recorded, and is sometimes implicit in the structure of the database or the typing of the attributes. We categorize the types of metadata that we have encountered in our work with machine learning applications in agriculture, and describe a first generation tool that we have built to aid in the recording and use of metadata in database mining.