A method for reducing unsafe concept associations in CLIP while preserving useful vision-language behavior.
Nov 27, 2023
