List your company

Safety Alignment on WildJailbreak

15,768.31Trainable parameters (M)Full FT
-630.59723,626.81147,884.2212,141.6286Feb 18, 2026

Evaluation results

MethodLinks
Full FTTarget Model=Qwen3-14B2026.0215,768.31
Full FTTarget Model=Qwen2.5-14B2026.0214,770.03
Full FTTarget Model=Phi-42026.0214,660
Full FTTarget Model=gemma-7b-it2026.028,537.68
Full FTTarget Model=Average2026.027,616.48
Full FTTarget Model=Qwen2.5-7B2026.027,615.61
Full FTTarget Model=Qwen3-4B2026.024,022.46
Full FTTarget Model=Phi-4-mini2026.023,836
Full FTTarget Model=Llama-3.2-3B2026.023,212.75
Full FTTarget Model=gemma-2b-it2026.022,506.17
Full FTTarget Model=Llama-3.2-1B2026.021,235.81
Circuit BreakerTarget Model=Phi-42026.0268.4
Circuit BreakerTarget Model=Qwen2.5-14B2026.0251.6
Circuit BreakerTarget Model=gemma-7b-it2026.0237.5
Circuit BreakerTarget Model=Qwen3-14B2026.0231.52
Circuit BreakerTarget Model=Qwen2.5-7B2026.0230.27
Circuit BreakerTarget Model=Average2026.0229.28
Circuit BreakerTarget Model=Qwen3-4B2026.0219.26
LoRATarget Model=Qwen2.5-14B2026.0218.81
Circuit BreakerTarget Model=Llama-3.2-3B2026.0218.23
Circuit BreakerTarget Model=gemma-2b-it2026.0214.16
Circuit BreakerTarget Model=Phi-4-mini2026.0213.36
Circuit BreakerTarget Model=Llama-3.2-1B2026.028.45
LoRATarget Model=gemma-2b-it2026.026.63
LoRATarget Model=Llama-3.2-3B2026.025.04
LoRATarget Model=Average2026.024.07
LoRATarget Model=Llama-3.2-1B2026.023.27
LoRATarget Model=Qwen3-14B2026.021.8
LoRATarget Model=Phi-4-mini2026.021.62
LoRATarget Model=gemma-7b-it2026.021.54
NeSTTarget Model=Qwen2.5-14B2026.021.01
LoRATarget Model=Qwen3-4B2026.020.89
NeSTTarget Model=Qwen3-14B2026.020.87
LoRATarget Model=Phi-42026.020.62
LoRATarget Model=Qwen2.5-7B2026.020.45
NeSTTarget Model=Qwen2.5-7B2026.020.45
NeSTTarget Model=Average2026.020.44
NeSTTarget Model=Phi-42026.020.42
NeSTTarget Model=Qwen3-4B2026.020.38
NeSTTarget Model=Llama-3.2-3B2026.020.36
NeSTTarget Model=gemma-7b-it2026.020.36
NeSTTarget Model=Phi-4-mini2026.020.21
NeSTTarget Model=gemma-2b-it2026.020.17
NeSTTarget Model=Llama-3.2-1B2026.020.13