List your company

Machine Unlearning on WMDP-chem 1.0 (test)

0.56BF16Zero-shot
0.219920.308210.39650.48479May 14, 2026

Evaluation results

MethodLinks
Zero-shotModel=Qwen-3-8B2026.050.560.560—0.741
Zero-shotModel=Qwen-3-4B2026.050.5410.5410—0.657
NPOModel=Qwen-3-4B2026.050.5210.5140.0070.5180.621
SimNPOModel=Qwen-3-4B2026.050.5170.5270.010.5110.628
GU+SimNPOModel=Qwen-3-4B2026.050.5140.5210.0070.5090.628
Zero-shotModel=Qwen-2.5-4B2026.050.5030.5030—0.61
Global GAModel=Llama-3.1-8B2026.050.4930.4730.020.4910.55
NPOModel=Qwen-2.5-4B2026.050.4830.4780.0050.4910.584
GU+SimNPOModel=Qwen-2.5-4B2026.050.4750.490.0150.4810.594
SimNPOModel=Qwen-2.5-4B2026.050.4750.4930.0180.4770.593
NPOModel=Llama-3.2-3B2026.050.4420.4250.0170.440.505
SimNPOModel=Llama-3.2-3B2026.050.4420.3930.0480.4490.505
Zero-shotModel=Llama-3.2-3B2026.050.4370.4370—0.538
GU+SimNPOModel=Llama-3.2-3B2026.050.4370.3950.0420.4450.503
Surgical GAModel=Llama-3.1-8B2026.050.4270.4230.0030.4260.525
MANSUModel=Qwen-3-4B2026.050.3340.3040.030.4340.641
MANSUModel=Llama-3.1-8B2026.050.3330.3070.0270.3980.584
Surgical GAModel=Qwen-3-8B2026.050.3130.3170.0030.3980.557
MANSUModel=Qwen-2.5-4B2026.050.3110.2840.0270.4040.594
MANSUModel=Qwen-3-8B2026.050.3070.2740.0330.3640.714
MANSUModel=Llama-3.2-3B2026.050.2740.2510.0230.3680.518
GU+SimNPOModel=Llama-3.1-8B2026.050.2730.27300.2310.23
NPOModel=Llama-3.1-8B2026.050.2530.2270.0270.2690.538
SimNPOModel=Qwen-3-8B2026.050.240.2770.0370.2590.405
Global GAModel=Qwen-3-8B2026.050.2370.2930.0570.2690.365
NPOModel=Qwen-3-8B2026.050.2370.23700.2960.515
SimNPOModel=Llama-3.1-8B2026.050.2330.23300.2410.195
GU+SimNPOModel=Qwen-3-8B2026.050.2330.23300.250.328