Safety Evaluation on Safety Benchmarks (Aggregate) and WildJailbreak (test)
95.7Safety ScoreReward shaping
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Reward shapingModel Family=Tulu 3 70B, Setting=λ = −22026.06 | 95.7 | 0.3 | 85.6 | 63.1 | |
| Reward shapingModel Family=OLMo 3 7B, Setting=λ = −52026.06 | 94.7 | 1.4 | 80.4 | 64 | |
| Reward shapingModel Family=OLMo 3.1 32B Instruct, Setting=λ = −22026.06 | 92.4 | 3.4 | 96.4 | 68.9 | |
| Reward shapingModel Family=Llama 3.1 8B, Setting=λ = −52026.06 | 91.7 | 4 | 84.4 | 42.5 | |
| Example filteringModel Family=Tulu 3 70B, Setting=drop 5%2026.06 | 91.7 | 5 | 94 | 62.2 | |
| Top-frac. steeringModel Family=Llama 3.1 8B, Setting=α = −5, f = 10%2026.06 | 91.4 | 5.9 | 82 | 42.8 | |
| SFTModel Family=Tulu 3 70B, Setting=–2026.06 | 91.1 | 7.4 | 88.4 | 50.7 | |
| Example filteringModel Family=OLMo 3.1 32B Instruct, Setting=drop 5%2026.06 | 90.8 | 5.1 | 97.6 | 68.8 | |
| SFTModel Family=OLMo 3.1 32B Instruct, Setting=–2026.06 | 90.4 | 9 | 90 | 58.8 | |
| Example filteringModel Family=OLMo 3 7B, Setting=drop 5%2026.06 | 90.2 | 5.8 | 93.2 | 63.3 | |
| Top-frac. steeringModel Family=OLMo 3 7B, Setting=α = −20, f = 3%2026.06 | 88.1 | 9.4 | 91.6 | 62.7 | |
| SFTModel Family=OLMo 3 7B, Setting=–2026.06 | 87.4 | 12 | 89.6 | 54.1 | |
| Explain-away steeringModel Family=Tulu 3 70B, Setting=α = −10, f = 5%2026.06 | 86.4 | 8.8 | 98.8 | 62.4 | |
| Top-frac. steeringModel Family=OLMo 3.1 32B Instruct, Setting=α = −20, f = 5%2026.06 | 85.1 | 11.6 | 99.6 | 70.1 | |
| SFTModel Family=Llama 3.1 8B, Setting=–2026.06 | 84.9 | 15.7 | 88.8 | 40.4 | |
| Example filteringModel Family=Llama 3.1 8B, Setting=drop 5%2026.06 | 83.5 | 15.6 | 96 | 42.9 | |
| DPOModel Family=OLMo 3 7B, Setting=–2026.06 | 83 | 13.2 | 98 | 62.9 | |
| DPOModel Family=OLMo 3.1 32B Instruct, Setting=–2026.06 | 81.5 | 14.3 | 100 | 69.8 | |
| Top-frac. steeringModel Family=Tulu 3 70B, Setting=α = −10, f = 5%2026.06 | 77.9 | 17.9 | 98.4 | 62.9 | |
| DPO reproModel Family=Tulu 3 70B, Setting=–2026.06 | 77.5 | 19.7 | 97.6 | 62.3 | |
| DPOModel Family=Llama 3.1 8B, Setting=–2026.06 | 75.8 | 21.7 | 98.8 | 43.4 |