Look at the “dataset” column: CLIP was trained on 400m images, UForm on 4m.