Safety Alignment Breaking Prevention on HarmBench
0Harmful Score (%)TokenBuncher
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| TokenBuncherModel=Qwen2.5-3B-Instruct, Harmful-RL version=w/o-attack2025.08 | 0 | 1.29 | |
| TokenBuncherModel=Qwen2.5-3B-Instruct, Harmful-RL version=RLOO2025.08 | 0 | 0.76 | |
| TokenBuncherModel=Qwen2.5-3B-Instruct, Harmful-RL version=REIN++2025.08 | 0 | 2.26 | |
| TokenBuncherModel=Qwen2.5-7B-Instruct, Harmful-RL version=w/o-attack2025.08 | 0 | 0.97 | |
| TokenBuncherModel=Qwen2.5-7B-Instruct, Harmful-RL version=RLOO2025.08 | 0 | 1.18 | |
| TokenBuncherModel=Mistral-8B-Instruct, Harmful-RL version=w/o-attack2025.08 | 0 | 0.26 | |
| TokenBuncherModel=Mistral-8B-Instruct, Harmful-RL version=PPO2025.08 | 0 | 0.43 | |
| TokenBuncherModel=Mistral-8B-Instruct, Harmful-RL version=RLOO2025.08 | 0 | 1.04 | |
| CTRAPModel=Qwen2.5-3B-Instruct, Harmful-RL version=w/o-attack2025.08 | 0.25 | 1.22 | |
| TokenBuncherModel=Qwen2.5-7B-Instruct, Harmful-RL version=REIN++2025.08 | 0.5 | 0.64 | |
| TokenBuncherModel=Qwen2.5-3B-Instruct, Harmful-RL version=PPO2025.08 | 0.75 | 2.62 | |
| TokenBuncherModel=Qwen2.5-3B-Instruct, Harmful-RL version=GRPO2025.08 | 0.75 | 2.44 | |
| TokenBuncherModel=Qwen2.5-7B-Instruct, Harmful-RL version=GRPO2025.08 | 0.75 | 0.59 | |
| TokenBuncherModel=Mistral-8B-Instruct, Harmful-RL version=GRPO2025.08 | 0.75 | 0.32 | |
| CTRAPModel=Qwen2.5-7B-Instruct, Harmful-RL version=w/o-attack2025.08 | 1 | 1.58 | |
| TokenBuncherModel=Mistral-8B-Instruct, Harmful-RL version=REIN++2025.08 | 1 | 0.43 | |
| TokenBuncherModel=Qwen2.5-7B-Instruct, Harmful-RL version=PPO2025.08 | 1.25 | 0.56 | |
| RepNoiseModel=Qwen2.5-3B-Instruct, Harmful-RL version=w/o-attack2025.08 | 3 | 2 | |
| CTRAPModel=Mistral-8B-Instruct, Harmful-RL version=w/o-attack2025.08 | 4 | 2.87 | |
| RepNoiseModel=Qwen2.5-7B-Instruct, Harmful-RL version=w/o-attack2025.08 | 6 | 1.13 | |
| BaseModel=Qwen2.5-7B-Instruct, Harmful-RL version=w/o-attack2025.08 | 7.75 | 4.2 | |
| RepNoiseModel=Mistral-8B-Instruct, Harmful-RL version=w/o-attack2025.08 | 7.75 | 0.68 | |
| BaseModel=Qwen2.5-3B-Instruct, Harmful-RL version=w/o-attack2025.08 | 13.75 | 4.04 | |
| BaseModel=Mistral-8B-Instruct, Harmful-RL version=w/o-attack2025.08 | 21.5 | 6.24 | |
| CTRAPModel=Qwen2.5-7B-Instruct, Harmful-RL version=RLOO2025.08 | 26.75 | 5.66 | |
| CTRAPModel=Qwen2.5-3B-Instruct, Harmful-RL version=RLOO2025.08 | 28.75 | 5.99 | |
| CTRAPModel=Mistral-8B-Instruct, Harmful-RL version=RLOO2025.08 | 28.75 | 8.67 | |
| CTRAPModel=Qwen2.5-3B-Instruct, Harmful-RL version=GRPO2025.08 | 33.5 | 6.38 | |
| CTRAPModel=Mistral-8B-Instruct, Harmful-RL version=GRPO2025.08 | 33.5 | 8.44 | |
| CTRAPModel=Qwen2.5-7B-Instruct, Harmful-RL version=GRPO2025.08 | 34.5 | 6.04 | |
| CTRAPModel=Qwen2.5-7B-Instruct, Harmful-RL version=REIN++2025.08 | 37.5 | 6.01 | |
| CTRAPModel=Qwen2.5-3B-Instruct, Harmful-RL version=PPO2025.08 | 40 | 7.6 | |
| CTRAPModel=Mistral-8B-Instruct, Harmful-RL version=PPO2025.08 | 40 | 6.86 | |
| RepNoiseModel=Qwen2.5-3B-Instruct, Harmful-RL version=RLOO2025.08 | 41 | 8.23 | |
| CTRAPModel=Qwen2.5-3B-Instruct, Harmful-RL version=REIN++2025.08 | 42 | 6.52 | |
| CTRAPModel=Mistral-8B-Instruct, Harmful-RL version=REIN++2025.08 | 42 | 8.3 | |
| RepNoiseModel=Mistral-8B-Instruct, Harmful-RL version=PPO2025.08 | 42.25 | 6.12 | |
| BaseModel=Qwen2.5-7B-Instruct, Harmful-RL version=PPO2025.08 | 42.75 | 9.13 | |
| CTRAPModel=Qwen2.5-7B-Instruct, Harmful-RL version=PPO2025.08 | 43 | 7.3 | |
| RepNoiseModel=Qwen2.5-7B-Instruct, Harmful-RL version=RLOO2025.08 | 43.75 | 8.7 | |
| BaseModel=Qwen2.5-7B-Instruct, Harmful-RL version=REIN++2025.08 | 44.25 | 9.08 | |
| BaseModel=Qwen2.5-7B-Instruct, Harmful-RL version=GRPO2025.08 | 45 | 9.04 | |
| BaseModel=Qwen2.5-7B-Instruct, Harmful-RL version=RLOO2025.08 | 45.5 | 8.96 | |
| BaseModel=Qwen2.5-3B-Instruct, Harmful-RL version=GRPO2025.08 | 45.75 | 8.83 | |
| BaseModel=Qwen2.5-3B-Instruct, Harmful-RL version=RLOO2025.08 | 46.5 | 8.78 | |
| RepNoiseModel=Mistral-8B-Instruct, Harmful-RL version=RLOO2025.08 | 46.5 | 9.24 | |
| BaseModel=Qwen2.5-3B-Instruct, Harmful-RL version=REIN++2025.08 | 46.75 | 8.99 | |
| RepNoiseModel=Qwen2.5-7B-Instruct, Harmful-RL version=PPO2025.08 | 46.75 | 8.66 | |
| BaseModel=Qwen2.5-3B-Instruct, Harmful-RL version=PPO2025.08 | 47.75 | 8.48 | |
| RepNoiseModel=Qwen2.5-3B-Instruct, Harmful-RL version=PPO2025.08 | 48.5 | 8.53 | |
| RepNoiseModel=Qwen2.5-7B-Instruct, Harmful-RL version=GRPO2025.08 | 48.5 | 8.98 | |
| RepNoiseModel=Qwen2.5-7B-Instruct, Harmful-RL version=REIN++2025.08 | 49 | 9.09 | |
| RepNoiseModel=Qwen2.5-3B-Instruct, Harmful-RL version=GRPO2025.08 | 49.75 | 8.35 | |
| RepNoiseModel=Mistral-8B-Instruct, Harmful-RL version=GRPO2025.08 | 50.5 | 9.03 | |
| BaseModel=Mistral-8B-Instruct, Harmful-RL version=RLOO2025.08 | 50.75 | 9.51 | |
| BaseModel=Mistral-8B-Instruct, Harmful-RL version=REIN++2025.08 | 51.25 | 9.32 | |
| BaseModel=Mistral-8B-Instruct, Harmful-RL version=PPO2025.08 | 52 | 8.54 | |
| RepNoiseModel=Mistral-8B-Instruct, Harmful-RL version=REIN++2025.08 | 53.25 | 5.99 | |
| BaseModel=Mistral-8B-Instruct, Harmful-RL version=GRPO2025.08 | 56.5 | 8.84 | |
| RepNoiseModel=Qwen2.5-3B-Instruct, Harmful-RL version=REIN++2025.08 | 65 | 6.86 |