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

Knowledge Unlearning on WMDP cyber

47.21Accuracy

Original model

23.071629.338335.60541.8717Jan 7, 2026Jan 24, 2026Feb 10, 2026Feb 27, 2026Mar 16, 2026Apr 2, 2026Apr 20, 2026
Updated 1mo ago

Evaluation Results

MethodLinks
2026.01
47.21
2026.01
46.9
2026.01
46.3
2026.01
46.25
2026.03
45.3
2026.03
45.3
2026.01
45.29
2026.03
44.3
2026.03
43.6
2026.03
43.1
2026.03
41.3
2026.01
38.2
2026.03
37
2026.04
35.78
2026.03
34.4
2026.01
33.87
2026.03
33.1
2026.03
32.1
2026.03
32.1
2026.01
31.58
2026.04
31.25
2026.04
30.95
2026.01
30.55
2026.04
30.05
2026.04
29.94
2026.04
29.89
2026.04
29.79
2026.04
29.79
2026.03
28.8
2026.04
28.64
2026.01
28.58
2026.01
27.98
2026.03
27.3
2026.01
27.22
2026.03
27.2
2026.01
27.13
2026.03
26.8
2026.03
26.6
2026.03
26.6
2026.01
26.52
2026.01
26.32
2026.03
26.3
2026.01
25.72
2026.03
25.6
2026.03
24.7
2026.03
24
2026.03
24