I have a quick question about the metric you reported. Does FWT in your paper refer to
the epoch-averaged success rate over {0, 5, … , 50} for each new task (as in LIBERO), or
the best success rate (maximum over epochs) for each new task?
Thanks a lot!