1 paper · 1 filter
Dylan Sam, Sachin Goyal, Pratyush Maini +2
Prior work has shown that safety interventions applied during pretraining, such as removing and rephrasing harmful content, can substantially improve the robustness of the resultin…