I agree the model is not transparent (I think that's what you meant, rather than opaque)
I agree it's hardly possible to grasp what the values in a model stand for
what I don't really see is that the training set offers better insights into the working of the model, especially when it comes to massive sets.
one might try to come up with theories, but in the end, whatever "understanding" one might infer from the training set is either equivalent and isomorphic to the model, or a misrepresentation of the model.
the former case would suggest that the model is its own source code, hard to grasp as it is
the latter case wouldn't really give you an understanding of the model

as for retraining, my limited knowledge of the subject matter suggests that additional training can bring the model to work as desired, regardless of the training that led to the initial model. it may be challenging to gain confidence that the further-trained model has no traces of responses one would like to suppress, but that is also the case in retraining from scratch, given the wild inferences and conclusions that such models may make from irrelevant details in the training sets. I'm curious as to any solid evidence you might have that might challenge this understanding
thanks for engaging!