Harmfulness Evaluation on HarmBench
1.12Harmful Response RatioBase
Evaluation Results
| Method | Links | |
|---|---|---|
| Base2025.10 | 1.12 | |
| Base2025.10 | 1.12 | |
| SFT2025.10 | 4.79 | |
| SFT2025.10 | 4.79 | |
| RLVR2025.10 | 4.96 | |
| RLVR2025.10 | 4.96 | |
| STAR-1Model Size=8B2026.01 | 21.26 | |
| THINKSAFEModel Size=8B2026.01 | 27.08 | |
| DirectRefusalModel Size=8B2026.01 | 32.39 | |
| SafeKeyModel Size=8B2026.01 | 32.72 | |
| THINKSAFEModel Size=7B2026.01 | 40.2 | |
| SafeChainModel Size=8B2026.01 | 44.52 | |
| SafeKeyModel Size=7B2026.01 | 45.35 | |
| SafePathModel Size=8B2026.01 | 47.51 | |
| SafeChainModel Size=7B2026.01 | 51 | |
| DirectRefusalModel Size=7B2026.01 | 52.33 | |
| InitialModel Size=8B2026.01 | 52.33 | |
| THINKSAFEModel Size=1.5B2026.01 | 52.99 | |
| STAR-1Model Size=7B2026.01 | 52.99 | |
| SafePathModel Size=7B2026.01 | 55.15 | |
| InitialModel Size=7B2026.01 | 56.98 | |
| SafeKeyModel Size=1.5B2026.01 | 58.8 | |
| SafeChainModel Size=1.5B2026.01 | 59.3 | |
| STAR-1Model Size=1.5B2026.01 | 62.79 | |
| SafePathModel Size=1.5B2026.01 | 65.28 | |
| DirectRefusalModel Size=1.5B2026.01 | 66.11 | |
| InitialModel Size=1.5B2026.01 | 67.28 |