As I understand things (not a user) abliteration has been superceded by actively monitoring the model state during the run and steering specific "negative" directions as they arise. It's both more reliable and does less damage.