Training data is a build dependency for the weights. You cannot realistically get the same weights without the same training data.
You mean the same source code? Because... I agree. That's why it's important for the source to be open. Both in the context of software and language models.