Strictly speaking, you could train a model which does not contain the original source text (just the underlying language structure and work tokens), and generates ASCII strings which are consistent with the underlying generative model, that are also always valid code. I expect to see code generator models that explicitly generate valid code as part of their generalization capability.