Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Safety Unlearning on BeaverTails

0Harmful Response Probability

RepSelect

-0.8084.64610.115.554Jun 15, 2026
Updated 1mo ago

Evaluation Results

MethodLinks
2026.06
0
2026.06
0
2026.06
0
2026.06
0
2026.06
0
2026.06
0.3
2026.06
0.4
2026.06
0.8
2026.06
0.9
2026.06
1
2026.06
1.3
2026.06
1.5
2026.06
2
2026.06
2.7
2026.06
3
2026.06
3.1
2026.06
3.1
2026.06
3.6
2026.06
4.6
2026.06
7.3
2026.06
7.8
2026.06
7.9
2026.06
8.5
2026.06
11.9
2026.06
12.8
2026.06
14.4
2026.06
14.5
2026.06
14.5
2026.06
14.8
2026.06
14.9
2026.06
15.1
2026.06
15.6
2026.06
15.6
2026.06
16.4
2026.06
17.6
2026.06
17.8
2026.06
17.8
2026.06
18.6
2026.06
19.9
2026.06
20
2026.06
20.2
2026.06
20.2