What do the concept labels represent?
Thanks for releasing this very cool resource!
I just read through the blog post and the dataset card and wasn't quite sure what the concept ids in this dataset represent -- are they the output of the concepts annotated by the mistral model (Stage 2) or from the concept annotator model (Stage 3)?
It's from a concept annotator model that was trained with the mistral annotations. Doing mistral annotations over the whole dataset would have been prohibitively expensive in terms of GPU hours, and also not given unified categories--If Mistral sometimes said "USA", and other times said "U.S.A.", we want to capture those pieces of text as being tagged with the same concept, so we need a fixed atlas of possible concepts that the concept annotator model outputs.