problems about poverty worst r about fish HOT 1 OPEN

Hi -- thanks for bringing this to my attention! I am working on a fix for this. If you are interested, I think the difference between performance can be explained as follows:

Our work was developed on WILDS v1.0, where the metric of interest for Poverty is average r instead of worst r. You can compare the current (v2.0) hyperparameter configuration of wilds vs. ours and notice some differences (for e.g. they use a step lr scheduler whereas we don't, and they use 8 groups per batch whereas we use random batches). I believe that once we implement the same hyperparameters as the current version of WILDS benchmark for Poverty we will observe similar performance with ERM using our implementation.

Stay tuned!!

from fish.