Could you provide class-conditioned VAR training with ImageNet? And we have a question: Is the model trained only on the train set?