Refusal behavior defense on WizardLM (test) benchmark leaderboard · SOTA2 Research

Refusal behavior defense on WizardLM (test)

90.4BadNet CACCInst_clean

Evaluation Results

Method	Links
Inst_clean2026.01		90.4	2.8	90.4	8.7	90.4	3.2	90.4	1.4
Fine-mixing2026.01		89	3.2	87.6	73.4	83.5	4.1	88.1	0.5
Clean-FFT2026.01		85.3	16.1	86.7	87.2	78.4	83	89.9	1.4
MB-Defense2026.01		81.2	0	88.9	1.8	78	0.9	76.6	0.5
Inst_atk2026.01		70.2	42.2	87.6	88.5	78	85.8	58.3	35.8
ONION2026.01		69.3	60.6	85	85.8	76	93.6	59.4	55
Inst_clean2026.01		57.8	3.2	57.8	33	57.8	5.5	57.8	0.9
MB-Defense2026.01		55	1.8	53.2	1.8	52.8	3.7	48.2	2.3
Fine-mixing2026.01		52.3	69.3	51.8	82.1	50	83	51.4	27.5
Clean-FFT2026.01		50.5	87.6	52.3	86.9	43.1	85.8	34.9	50.5
Inst_atk2026.01		48.6	87.6	48.6	87.2	45	86.7	16.5	87
ONION2026.01		46.8	45	48.6	88.5	48.2	85.8	13.4	82.6