Tag
This paper assesses the prevalence of extremist speech in the training data of large language models, focusing on the Dolma corpus, and discusses implications for data curation and model safety.