#Olmo-1B

New research shows that including toxic data in LLM pretraining improves the model's ability to be detoxified and controlled, leading to safer and more robust language models.