Is your feature request related to a problem? Please describe.
The inverse triples usage is currently somewhat inconsistent, e.g., we expected the training instances to contain inverse triples, but do require evaluation triples to not contain them.

Describe the solution you'd like
Encapsulate inverse triples logic into Model class. This also would be coherent with our description in the resource and benchmarking paper, where we explain why using inverse relations is a modelling decision rather than a simple data augmentation technique.