Previous works on finetuning safety largely target misuse-related finetuning attacks that make models comply with harmful requests (‘jailbreak finetuning’ 17). We ran head-to-head evaluations between ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results