I guess it belongs in the Front-end.
Data reduction is only useful when different users open your document(s) and need to see different subsets of your data. But AFAIK there is only one user that reads Stage-3 Datamodels, i.e. the QDS service account. If that user always gets the exact same data reduction, then why not simply change your script to load fewer records to begin with?
In addition to that, a load script run often undoes the effect of a data reduction, because the first thing that happens during a document reload (like in Stage 3) is a wipe of the entire data model. Whatever comes after is not influenced by a data reduction or section access (except when doing a BINARY load).