A less generic middle ground could be to have generic dataset for each library then just specify the reader / writer methods I guess. The other route is to make implementation of dataset as simple as providing load / save functions with the right signature. Will try to create a feature request in the next few days.