a question about the ood dataset

[FayeXXX]

Thanks for your excellent work! I was reviewing the implementation in train_eval_util.py and noticed an interesting design choice:
https://github.com/deeplearning-wisc/MCM/blob/ea7130f851e7d462cacd21f0e87a127705700bd9/utils/train_eval_util.py#L141
• ImageNet10 uses the training split
• ImageNet20 uses the validation split

Could you please share the rationale behind this dataset partitioning strategy?