You are exactly right! There does not seem to be much out there. I usually do hash or otherwise transform real-world (e.g., https://github.com/awesomedata/awesome-public-datasets) data to retain basic properties like skewness utf8 well-formedness, cardinality, and such. Like many others, I often use the Enron eMail dataset, NYC Taxi, or Twitter data (for coherent use of Unicode). And then scale it either by combining attributes randomly or otherwise augmenting it. Earlier in my career, I wrote a regular expression inversion data generator; it quickly spiraled out of control complexity-wise ;-).