iNat2021 is a large image dataset containing more than 2.7 million images from 10,000 different species collected and annotated by scientists.

To make the dataset more accessible, the authors also created a "mini" training dataset. There are 50 training images per species, for a total of 500,000 training images. There are 10 validation images per species, for a total of 100,000 validation images. There are 500,000 test images in total. Each species has at least 152 images in the iNat2021 training set.