Make evaluation configurable

[tsterbak]

ATM evaluation only supports positive/negative as rating. But we could imagine rating on a scale, multilabel rating or something else. This should be configurable at dataset creation time.