More distillation details about training datasets

[liuxy1103]

I would like to know what distillation algorithm you used specifically, which small model was trained on which data set, and whether it still has the ability of "emergence", that is, the ability of zero-shot segmentation