Distribution of the training dataset #7
Description
Could you publicly share the distribution of your training datasets?
For example, how much data did you use for Widget Captioning, UI RefExp, ShowUI-web, and OmniAct?
For RefExp, did you use dev.tfrecord, test.tfrecord, and train.tfrecord, or only train.tfrecord?
For OmniAct, did you use train.json and test.json, or only train.tfrecord? Also, how did you obtain the bounding boxes (bbox)? I tried filtering by checking whether a point falls inside a bbox, but only about 900 cases succeeded. Could you provide some details?
Additionally, with the same configuration (8 × A100 80G), my training time far exceeds one day. I believe this is related to your training data volume. Thank you for your response and for open-sourcing.