List your company

Safety on WildJailbreak

17.6Harmful Response RatioSTAR-1
16.17825.776535.37544.9735Jan 30, 2026

Evaluation results

MethodLinks
STAR-1Model Size=8B2026.0117.6—
THINKSAFEModel Size=8B2026.0121.15—
SafeKeyModel Size=8B2026.0130.85—
DirectRefusalModel Size=8B2026.0132.6—
THINKSAFEModel Size=7B2026.0135.4—
THINKSAFEModel Size=1.5B2026.0140.5—
SafeChainModel Size=8B2026.0142.45—
SafeKeyModel Size=7B2026.0145.75—
SafeChainModel Size=7B2026.0145.85—
SafeChainModel Size=1.5B2026.0146.95—
SafeKeyModel Size=1.5B2026.0147.65—
STAR-1Model Size=7B2026.0148.75—
STAR-1Model Size=1.5B2026.0149.65—
InitialModel Size=8B2026.0149.7—
DirectRefusalModel Size=7B2026.0150.2—
SafePathModel Size=8B2026.0150.2—
DirectRefusalModel Size=1.5B2026.0150.45—
InitialModel Size=1.5B2026.0151.55—
SafePathModel Size=1.5B2026.0151.8—
SafePathModel Size=7B2026.0152.65—
InitialModel Size=7B2026.0153.15—
DeepSeek-V3.2Mode=Non-Reasoning, Architecture=MoE, Total Params=671B, Activated Params=37B2026.01—76.8
DGR SFTAlignment Dataset=DirectRefusal, Backbone=s1.1-7B2026.02—91.2
DGR SFTAlignment Dataset=STAR-1, Backbone=s1.1-7B2026.02—97.2
DGR SFTAlignment Dataset=R1-ACT, Backbone=s1.1-7B2026.02—100
EXAONE 4.0Mode=Non-Reasoning, Architecture=Dense, Total Params=32B, Activated Params=32B2026.01—49.6
K-EXAONEMode=Non-Reasoning, Architecture=MoE, Total Params=236B, Activated Params=23B2026.01—91.6
Qwen2.5-7B-Instruct2026.02—56.8
Qwen3-235B-A22B Instruct-2507Architecture=MoE, Total Params=235B, Activated Params=22B2026.01—91.1
s1.1-7B2026.02—48.8
Vanilla SFTAlignment Dataset=DirectRefusal, Backbone=s1.1-7B2026.02—98.8
Vanilla SFTAlignment Dataset=STAR-1, Backbone=s1.1-7B2026.02—95.6
Vanilla SFTAlignment Dataset=R1-ACT, Backbone=s1.1-7B2026.02—100