about data processing

[starhiking]

Is the grounding data labeled before or after the data processor?

for example, an image needs padding to make width equal height, and the openai/clip-vit contains the do_center_crop. These operations will change the grounding location. But the labels in conversations remain the same. So I wonder if it will produce an impact on training.

[zyandtom]

hi, May I ask where this grounding dataset comes from? Also, does llava onevision have any special tokens for bboxes, or is there any preprocessing func for box? I didn't see it in the code.