is it the llm model weights or the training data that's important and confidential?