This costly curation process limits the size of datasets and hence hinders the scaling of trained models.