Cloud-specific languages are usually untyped, and no guarantees about the correctness of complex jobs can be statically obtained. Datasets too are usually untyped and no schema information is needed for their manipulation. In this paper we sketch a typing algorithm for JSON datasets. Our approach can be used to infer a succinct type from scratch for a collection of JSON objects, as well as to validate a dataset against a human-designed type and, if necessary, to adapt and improve this type.