The first iteration of our model was built off of this amazing public dataset:

https://www.gwern.net/Danbooru2020

Though now we have made our own :)