For some learners, we need to store information beyond the model object from the upstream package.

Therefore, we solve this by making the private $.train function of a learner return something along the lines of:

list(
  model = upstream::train(...),
  info = ...
)

this means that learner$model does now not return the 'native' model, but the named list from above.

We could add support for an active binding $native_model , that can be implemented by learners that do this and which would then return the actual upstream model.