That is not reproducible research. To reproduce the research, you need training data, source code and all the parameters.

With black-boxed generated models, there is no way to tell how they have been actually generated. It has no value for science to how to improve it further.