Holistic Evaluation of Language Models (HELM), a framework to increase the transparency of language models (https://arxiv.org/abs/2211.09110). This framework is also used to evaluate text-to-image models in Holistic Evaluation of Text-to-Image Models (HEIM) (https://arxiv.org/abs/2311.04287).
We currently use some Kaggle data for CivilComments, but we should use the more authoritative WILDS data (that also merge sparse categories in a sensible way; I guess the other religion category makes sense, but is a bit weird given judaism/hinduism/buddhism are major world religions, though just not that frequent in the data).
We should have a flag to toggle between, and once done, we should switch to WILDS as the default.
@rishibommasani, do we still want to support both, or should we just use the WILDS version? @ryanachi If you don't have the bandwidth, I can take this.
We currently use some Kaggle data for CivilComments, but we should use the more authoritative WILDS data (that also merge sparse categories in a sensible way; I guess the other religion category makes sense, but is a bit weird given judaism/hinduism/buddhism are major world religions, though just not that frequent in the data).
We should have a flag to toggle between, and once done, we should switch to WILDS as the default.