We trained six different classifiers to detect and remove CBRN information from training data. The best and most efficient results were from a classifier that used a small model from the Claude 3 Sonnet series to flag the harmful data.

→ View original post on X — @anthropicai