Our new X account is live! Follow @wizwand_team for updates

Safety Evaluation on Harmbench

0.06Harmbench Score

NPO

Updated 2d ago

Evaluation Results

Method	Links
NPO 2025.04		0.06
Safety Prompting 2025.04		0.31
REPBEND 2025.04		0.31
REPBEND 2025.04		1.56
SFT 2025.04		1.56
TA 2025.04		1.87
Safety 2026.01		2
LED 2026.01		2
ReasonAny 2026.01		2
NPO 2025.04		2.19
DARE 2026.01		3
ReasonAny 2026.01		3
DPO 2025.04		3.13
TA 2025.04		3.75
RMU 2025.04		3.75
CB 2025.04		3.75
Qwen2.5-32B-Instruct (Safety) 2026.01		4
ReasonAny 2026.01		4
Linear 2026.01		4
DARE 2026.01		4
SFT 2025.04		4.69
NPO (W/O safe set) 2025.04		5.63
R2D2* 2025.04		5.63
Linear 2026.01		6
DARE 2026.01		6
Qwen2.5-14B-Instruct (Safety) 2026.01		6
Task Arithmetic 2026.01		9
RMU 2025.04		9.37
Surgery 2026.02		9.5
Linear 2026.01		10
Task Arithmetic 2026.01		11
NPO (W/O safe set) 2025.04		11.25
WHP 2025.04		12.5
FuseLLM 2026.01		13
CB* 2025.04		13.44
CB* 2025.04		13.44
Task Arithmetic 2026.01		16
CB 2025.04		16.25
WHP (W/O safe set) 2025.04		18.75
Safety Prompting 2025.04		20
TIES 2026.01		20
Lisa 2026.02		20.75
DeepSeek-R1-Distill-Qwen-14B (Reasoning) 2026.01		21
ConstrainedSFT 2026.02		21.25
Original 2025.04		22.19
AsFT 2026.02		22.5
SafeGrad 2026.02		23.5
DSS 2026.02		24.5
DPO 2025.04		25.63
DeepSeek-R1-Distill-Qwen-32B (Reasoning) 2026.01		26
SFT 2026.02		28.75
FuseLLM 2026.01		29
SPARD 2026.02		30
Reasoning 2026.01		30
LED 2026.01		33
WHP 2025.04		33.13
DeepSeek-R1 2025.01		35
FuseLLM 2026.01		35
LED 2026.01		35
TIES 2026.01		46
WHP (W/O safe set) 2025.04		49.06
Original 2025.04		50.94
TIES 2026.01		56
DeepSeek-R1 2025.01		58
DeepSeek-V3 2025.01		67
GPT-4o (2024-05-13) 2025.01		72.7
Vanilla 2026.02		81.2
Qwen2.5 Instruct (72B) 2025.01		83
OmniGuard 2026.02		83.2
Claude-3.7-Sonnet 2025.01		83.3
o1 (2024-12-17) 2025.01		84
DeepSeek-R1 2025.01		89.3
Self-Reminder 2026.02		93.3
DeepSeek-V3 2025.01		96
DeepSeek-R1 2025.01		96.3
OmniSteer 2026.02		100