Datasets with data from multiple collections/institutions

[MichalTorma]

So we had this request popping up about publishing datasets that would contain records from different collections that are not yet published. The question is if we should endorse this.
The issue that may arise once the actual collections are published as a whole is that there will be duplicity. also in the case of the dataset containing data from multiple institutions (but extracted and revised by a researcher from one institution), citation would be accredited to the institution that published the dataset not the actual owner of the data.
GBIF on the other hand has data clustering feature that should be able to deal with this.
Any thoughts?