Also of interest might be cleanlab (https://github.com/cgnorthcutt/cleanlab), the open-source software we used for initially identifying potentially mislabeled data.