General Capability on MTBench
9.14MTBench ScoreREPBEND
Evaluation Results
| Method | Links | |
|---|---|---|
| REPBENDModel=Qwen2.5 14B Instruct2025.04 | 9.14 | |
| Original Weight ModelModel=Qwen2.5 14B Instruct2025.04 | 8.71 | |
| WHP (W/O safe set)Backbone=Llama3 8B Instruct2025.04 | 8.01 | |
| WHP (W/O safe set)Backbone=Mistral 7B Instruct-v0.22025.04 | 7.88 | |
| OriginalBackbone=Llama3 8B Instruct2025.04 | 7.84 | |
| NPOBackbone=Llama3 8B Instruct2025.04 | 7.79 | |
| CBBackbone=Llama3 8B Instruct2025.04 | 7.74 | |
| OriginalBackbone=Mistral 7B Instruct-v0.22025.04 | 7.72 | |
| CB*Backbone=Llama3 8B Instruct, publicly_available_safety_tuned=true2025.04 | 7.72 | |
| NPO (W/O safe set)Backbone=Llama3 8B Instruct2025.04 | 7.71 | |
| REPBENDBackbone=Llama3 8B Instruct2025.04 | 7.71 | |
| APL (RM)Model=Llama-3-8B2025.05 | 7.65 | |
| WHPBackbone=Llama3 8B Instruct2025.04 | 7.63 | |
| CAPOModel=Llama-3-8B2025.05 | 7.63 | |
| APLModel=Llama-3-8B2025.05 | 7.59 | |
| CBBackbone=Mistral 7B Instruct-v0.22025.04 | 7.58 | |
| DPOBackbone=Mistral 7B Instruct-v0.22025.04 | 7.57 | |
| CB*Backbone=Mistral 7B Instruct-v0.2, publicly_available_safety_tuned=true2025.04 | 7.57 | |
| APL (w/o Adv)Model=Llama-3-8B2025.05 | 7.56 | |
| DPOBackbone=Llama3 8B Instruct2025.04 | 7.54 | |
| DPOModel=Llama-3-8B2025.05 | 7.54 | |
| BaseModel=Llama-3-8B2025.05 | 7.53 | |
| Safety PromptingBackbone=Mistral 7B Instruct-v0.22025.04 | 7.5 | |
| REPBENDBackbone=Mistral 7B Instruct-v0.22025.04 | 7.5 | |
| REPBENDModel=Gemma2 2B Instruct2025.04 | 7.37 | |
| Original Weight ModelModel=Gemma2 2B Instruct2025.04 | 7.35 | |
| NPOBackbone=Mistral 7B Instruct-v0.22025.04 | 7.31 | |
| Safety PromptingBackbone=Llama3 8B Instruct2025.04 | 7.18 | |
| TABackbone=Llama3 8B Instruct2025.04 | 7 | |
| TABackbone=Mistral 7B Instruct-v0.22025.04 | 6.93 | |
| SFTBackbone=Llama3 8B Instruct2025.04 | 6.88 | |
| APL (RM)Model=Mistral-7B2025.05 | 6.88 | |
| DPOModel=Mistral-7B2025.05 | 6.85 | |
| BaseModel=Mistral-7B2025.05 | 6.78 | |
| CAPOModel=Mistral-7B2025.05 | 6.66 | |
| APLModel=Mistral-7B2025.05 | 6.59 | |
| APL (w/o Adv)Model=Mistral-7B2025.05 | 6.53 | |
| WHPBackbone=Mistral 7B Instruct-v0.22025.04 | 6.29 | |
| SFTBackbone=Mistral 7B Instruct-v0.22025.04 | 5.98 | |
| R2D2*Backbone=Mistral 7B Instruct-v0.2, publicly_available_safety_tuned=true2025.04 | 5.97 | |
| RMUBackbone=Mistral 7B Instruct-v0.22025.04 | 5.77 | |
| NPO (W/O safe set)Backbone=Mistral 7B Instruct-v0.22025.04 | 4.67 | |
| RMUBackbone=Llama3 8B Instruct2025.04 | 4.01 |