The RVL-CDIP-N corpus is from Larson et al. 2022. It consists of 1,002 out-of-distribution samples for evaluating model generalizability for models trained on RVL-CDIP. RVL-CDIP-N can be found here.