There are still things you can't do with an open-weight model without the training data, like modifying the architecture and training from scratch. That's different from true open-source code, where you can do anything the authors could do.