zlacker

The worry is about dataset shifting. Previously, the data were collected for a few hundreds thousands structures, now it is 200m. I think there could be doubts on distributions and how that could play a role in prediction accuracy.