Jailbreaking Safety Evaluation on Fortress
87.7Safety ScoreSelf-ReSET
Evaluation Results
| Method | Links | |
|---|---|---|
| Self-ReSETBase Model=DeepSeek-R1-Distill-Llama-8B2026.05 | 87.7 | |
| RECAPBackbone=DSLlama-8B2025.10 | 86.84 | |
| Self-ReSETBase Model=DeepSeek-R1-Distill-Qwen-7B2026.05 | 82.4 | |
| RECAPBase Model=DeepSeek-R1-Distill-Llama-8B2026.05 | 81.5 | |
| SEARModel Size=1.5B, Pre-fill attack protocol=true2026.06 | 81.5 | |
| RECAPBackbone=DSQwen-14B2025.10 | 80.17 | |
| STARBackbone=DSQwen-14B2025.10 | 79.05 | |
| STARBackbone=DSLlama-8B2025.10 | 78.49 | |
| Self-ReSETBase Model=Qwen3-8B2026.05 | 77.2 | |
| SFTBackbone=DSQwen-14B2025.10 | 74.54 | |
| SFTBackbone=DSLlama-8B2025.10 | 74.25 | |
| DAPOBase Model=DeepSeek-R1-Distill-Llama-8B2026.05 | 74.2 | |
| STAR-1Base Model=Qwen3-8B2026.05 | 72.9 | |
| STAR-1Base Model=DeepSeek-R1-Distill-Llama-8B2026.05 | 72.6 | |
| RECAPBase Model=DeepSeek-R1-Distill-Qwen-7B2026.05 | 69.5 | |
| STAR-1Base Model=DeepSeek-R1-Distill-Qwen-7B2026.05 | 69 | |
| DAPOBackbone=DSLlama-8B2025.10 | 68.86 | |
| DAPOBackbone=DSQwen-14B2025.10 | 67.85 | |
| DAPOBase Model=DeepSeek-R1-Distill-Qwen-7B2026.05 | 66.4 | |
| QWEN3 32BModel Size=32B, Pre-fill attack protocol=true2026.06 | 65.9 | |
| SafeChainBackbone=DSLlama-8B2025.10 | 65.56 | |
| QWEN3 14BModel Size=14B, Pre-fill attack protocol=true2026.06 | 65.5 | |
| QWEN3 8BModel Size=8B, Pre-fill attack protocol=true2026.06 | 63.8 | |
| SafeChainBackbone=DSQwen-14B2025.10 | 63.15 | |
| DAPOBase Model=Qwen3-8B2026.05 | 62.7 | |
| RECAPBase Model=Qwen3-8B2026.05 | 62.5 | |
| OriginalBackbone=DSLlama-8B2025.10 | 60.83 | |
| OriginalBackbone=DSQwen-14B2025.10 | 57.91 | |
| SafechainBase Model=DeepSeek-R1-Distill-Llama-8B2026.05 | 57.3 | |
| SafechainBase Model=Qwen3-8B2026.05 | 56 | |
| SafechainBase Model=DeepSeek-R1-Distill-Qwen-7B2026.05 | 53.5 | |
| BaseBase Model=DeepSeek-R1-Distill-Qwen-7B2026.05 | 49.6 | |
| BaseBase Model=Qwen3-8B2026.05 | 48.9 | |
| BaseBase Model=DeepSeek-R1-Distill-Llama-8B2026.05 | 48 |