Safety Evaluation on SafetyBench (accuracy)
89AccuracyInvThink SFT+RL
Evaluation Results
| Method | Links | |
|---|---|---|
| InvThink SFT+RLBackbone=Qwen-3-8B2025.10 | 89 | |
| InvThink SFTBackbone=Qwen-3-8B2025.10 | 87 | |
| InvThinkBackbone=Qwen-3-8B2025.10 | 85 | |
| General SFT+RLBackbone=Qwen-3-8B2025.10 | 85 | |
| Gemini-2.5 ProBackbone=Teacher Model2025.10 | 85 | |
| SafetyPromptBackbone=Qwen-3-8B2025.10 | 84 | |
| CoTBackbone=Qwen-3-8B2025.10 | 83 | |
| InvThink SFT+RLBackbone=Qwen-2.5-7B2025.10 | 82 | |
| General SFTBackbone=Qwen-3-8B2025.10 | 82 | |
| InvThink SFTBackbone=Qwen-2.5-7B2025.10 | 79 | |
| InvThink SFT+RLBackbone=Gemma-7B-it2025.10 | 77 | |
| General SFT+RLBackbone=Qwen-2.5-7B2025.10 | 77 | |
| ToTBackbone=Qwen-3-8B2025.10 | 77 | |
| InvThink SFTBackbone=Gemma-7B-it2025.10 | 76 | |
| CoTBackbone=Qwen-2.5-7B2025.10 | 76 | |
| InvThinkBackbone=Qwen-2.5-7B2025.10 | 76 | |
| General SFTBackbone=Qwen-2.5-7B2025.10 | 76 | |
| Zero-shotBackbone=Qwen-3-8B2025.10 | 76 | |
| SafetyPromptBackbone=Qwen-2.5-7B2025.10 | 75 | |
| General SFT+RLBackbone=Gemma-7B-it2025.10 | 74 | |
| InvThinkBackbone=Gemma-7B-it2025.10 | 73 | |
| Zero-shotBackbone=Qwen-2.5-7B2025.10 | 73 | |
| Zero-shotBackbone=Gemma-7B-it2025.10 | 72 | |
| General SFTBackbone=Gemma-7B-it2025.10 | 72 | |
| ToTBackbone=Qwen-2.5-7B2025.10 | 71 | |
| CoTBackbone=Gemma-7B-it2025.10 | 69 | |
| SafetyPromptBackbone=Gemma-7B-it2025.10 | 67 | |
| ToTBackbone=Gemma-7B-it2025.10 | 62 |