This way we could then return nice, human-understandable groups rightaway - not needing the tricky `report_subgroup_bias` function in utils.py.