I'd actually say that including the training data would be like providing the IDE/debugger/compiler rather than the model/checkpoint being analogous. If I hand you Signal's source code you can run it, use it, modify it, etc. All similar characteristics to what is provided here. What they didn't provide to us is how they created that code. You couldn't create that software from scratch by just having these and that's true for any open source project. But I wouldn't say training data is as good as an analogy to peering in the minds of engineers, because it is an important part to getting the final product and analyzing it.