I wonder, which of the scoring functions available in Modeller is
preferable to choose best models.
There isn't really a debate here - the answer is DOPE or SOAP.
The fact that molpdf only is selected as the default score of model
quality makes DOPE at little suspicious.
molpdf is the Modeller scoring function, so it is always output - we
can't build models without it. Modeller computes no assessment functions
by default. Note that an assessment function and a scoring function are
not the same thing, and are designed to solve different problems
(assessment shows how like a "normal protein" your model is; scoring
shows how like the template it is).
Restraint violations are an indication of problems with your inputs (the
alignment, typically). You shouldn't use them to rank models.