Thanks for reporting.
This sounds like a recent regression stemming from us keeping uploaded images in sequential order to make it easier for annotators to label sequences of video frames.
Let me look at the code and see what I can do to restore the randomness for the model while preserving the ordering for humans in the tool.