Safety Evaluation on HEx-PHI
97.2HEx-PHI ScoreID-LoRA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ID-LoRABase Model=Mistral-7B, # Params (%)=0.62%2026.02 | 97.2 | — | |
| HydraLoRABase Model=Mistral-7B, # Params (%)=1.22%2026.02 | 97 | — | |
| ID-LoRABase Model=LLaMA-3-8B, # Params (%)=0.56%2026.02 | 95.3 | — | |
| DoRABase Model=Mistral-7B, # Params (%)=1.16%2026.02 | 95.3 | — | |
| FFTBase Model=LLaMA-3-8B, # Params (%)=100%2026.02 | 94.8 | — | |
| HydraLoRABase Model=LLaMA-3-8B, # Params (%)=1.10%2026.02 | 94.2 | — | |
| FFTBase Model=Mistral-7B, # Params (%)=100%2026.02 | 94.1 | — | |
| TOSS-ProModel=Llama-3-8B-Instruct2026.03 | 93.79 | — | |
| DoRABase Model=LLaMA-3-8B, # Params (%)=1.05%2026.02 | 93.6 | — | |
| LoRABase Model=Mistral-7B, # Params (%)=1.15%2026.02 | 91.9 | — | |
| LoRABase Model=LLaMA-3-8B, # Params (%)=1.03%2026.02 | 91.6 | — | |
| TOSSModel=Llama-3-8B-Instruct2026.03 | 87.54 | — | |
| TOSS-ProModel=Llama-2-7B-Chat-hf2026.03 | 74.43 | — | |
| TOSSModel=Llama-2-7B-Chat-hf2026.03 | 69.9 | — | |
| SEALModel=Llama-3-8B-Instruct2026.03 | 68.83 | — | |
| RandomModel=Llama-3-8B-Instruct2026.03 | 65.03 | — | |
| SafeInstrModel=Llama-3-8B-Instruct2026.03 | 64.63 | — | |
| DSIRModel=Llama-3-8B-Instruct2026.03 | 60.83 | — | |
| DSIRModel=Llama-2-7B-Chat-hf2026.03 | 56.99 | — | |
| SafeInstrModel=Llama-2-7B-Chat-hf2026.03 | 51.34 | — | |
| SEALModel=Llama-2-7B-Chat-hf2026.03 | 50.33 | — | |
| Standard SFTModel=Llama-3-8B-Instruct2026.03 | 50 | — | |
| Standard SFTModel=Llama-2-7B-Chat-hf2026.03 | 50 | — | |
| RandomModel=Llama-2-7B-Chat-hf2026.03 | 49.16 | — | |
| OriginalModel=Llama-3.1-8B, # Params=-2025.12 | 4.09 | — | |
| OriginalModel=Gemma-2-9B, # Params=-2025.12 | 3.95 | — | |
| OriginalModel=Gemma-2-2B, # Params=-2025.12 | 3.63 | — | |
| LoRAModel=Gemma-2-9B, # Params=0.19%2025.12 | 1.54 | — | |
| DoRAModel=Gemma-2-9B, # Params=0.20%2025.12 | 1.48 | — | |
| LoRAModel=Gemma-2-2B, # Params=0.24%2025.12 | 1.46 | — | |
| FFTModel=Gemma-2-2B, # Params=100%2025.12 | 1.39 | — | |
| DoRAModel=Llama-3.1-8B, # Params=0.22%2025.12 | 1.36 | — | |
| PromptModel=Llama-3.1-8B, # Params=-2025.12 | 1.33 | — | |
| PromptModel=Gemma-2-2B, # Params=-2025.12 | 1.32 | — | |
| DoRAModel=Gemma-2-2B, # Params=0.25%2025.12 | 1.31 | — | |
| LoRAModel=Llama-3.1-8B, # Params=0.21%2025.12 | 1.27 | — | |
| PromptModel=Gemma-2-9B, # Params=-2025.12 | 1.26 | — | |
| FFTModel=Gemma-2-9B, # Params=100%2025.12 | 1.24 | — | |
| FFTModel=Llama-3.1-8B, # Params=100%2025.12 | 1.16 | — | |
| IT+RLModel=Llama-3.1-8B, # Params=100%2025.12 | 1.16 | — | |
| SAILSModel=Gemma-2-2B, # Params=0.24%2025.12 | 1.08 | — | |
| SAILSModel=Llama-3.1-8B, # Params=0.21%2025.12 | 1.03 | — | |
| IT+RLModel=Gemma-2-2B, # Params=100%2025.12 | 1.02 | — | |
| SAILSModel=Gemma-2-9B, # Params=0.19%2025.12 | 1.01 | — | |
| IT+RLModel=Gemma-2-9B, # Params=100%2025.12 | 1 | — | |
| SkeletonModel=Llama-2-13B2026.01 | 1 | — | |
| IDGAFModel=Llama-2-13B2026.01 | 1 | — | |
| AOAModel=Gemma-1.1-7B2026.01 | 1 | — | |
| BaseModel=Gemma-2-9B2026.01 | 1 | — | |
| AOAModel=Gemma-2-9B2026.01 | 1 | — | |
| SkeletonModel=Gemma-2-9B2026.01 | 1 | — | |
| SkeletonBackbone=GPT-oss-20B2026.01 | 1 | — | |
| BaseLLM Model=Gemma-2-27B2026.01 | 1 | — | |
| AOALLM Model=Gemma-2-27B2026.01 | 1 | — | |
| AOAModel=Llama-2-13B2026.01 | 0.997 | — | |
| SkeletonModel=Gemma-1.1-7B2026.01 | 0.997 | — | |
| IDGAFModel=Gemma-1.1-7B2026.01 | 0.997 | — | |
| FormalModel=Gemma-2-9B2026.01 | 0.997 | — | |
| IDGAFModel=Gemma-2-9B2026.01 | 0.997 | — | |
| FormalLLM Model=Gemma-2-27B2026.01 | 0.997 | — | |
| SkeletonLLM Model=Gemma-2-27B2026.01 | 0.9939 | — | |
| IDGAFLLM Model=Gemma-2-27B2026.01 | 0.9939 | — | |
| BaseModel=Gemma-2-2B2026.01 | 0.9939 | — | |
| BaseBackbone=GPT-oss-20B2026.01 | 0.9909 | — | |
| AOABackbone=GPT-oss-20B2026.01 | 0.9909 | — | |
| FormalModel=Gemma-1.1-7B2026.01 | 0.9879 | — | |
| FormalModel=Gemma-2-2B2026.01 | 0.9879 | — | |
| FormalModel=Llama-2-13B2026.01 | 0.9848 | — | |
| FormalBackbone=GPT-oss-20B2026.01 | 0.9848 | — | |
| BaseModel=Gemma-1.1-7B2026.01 | 0.9818 | — | |
| IDGAFBackbone=GPT-oss-20B2026.01 | 0.9818 | — | |
| GCGModel=Gemma-2-9B2026.01 | 0.9784 | — | |
| GCGLLM Model=Gemma-2-27B2026.01 | 0.9784 | — | |
| BaseModel=Llama-2-13B2026.01 | 0.9727 | — | |
| GCGModel=Gemma-1.1-7B2026.01 | 0.9722 | — | |
| Refusal SuppressionModel=Llama-2-13B2026.01 | 0.9697 | — | |
| GCGBackbone=GPT-oss-20B2026.01 | 0.9696 | — | |
| AOAModel=Gemma-2-2B2026.01 | 0.9606 | — | |
| IDGAFLLM Model=Qwen2.5-32B2026.01 | 0.9576 | — | |
| AOAModel=Llama-3.2-1B2026.01 | 0.9545 | — | |
| AOALLM Model=Llama-3.3-70B2026.01 | 0.9545 | — | |
| BaseLLM Model=Qwen2.5-32B2026.01 | 0.9545 | — | |
| Refusal SuppressionModel=Gemma-1.1-7B2026.01 | 0.9515 | — | |
| IDGAFModel=Llama-3.1-8B2026.01 | 0.9485 | — | |
| AOAModel=Qwen2-7B2026.01 | 0.9485 | — | |
| GCGLLM Model=Qwen2.5-32B2026.01 | 0.9483 | — | |
| AOAModel=Llama-3.1-8B2026.01 | 0.9455 | — | |
| IDGAFModel=Qwen3-4B-think2026.01 | 0.9424 | — | |
| AOALLM Model=Qwen2.5-32B2026.01 | 0.9394 | — | |
| FormalModel=Llama-3.1-8B2026.01 | 0.9394 | — | |
| GCGModel=Llama-2-13B2026.01 | 0.9364 | — | |
| IDGAFModel=Gemma-2-2B2026.01 | 0.9364 | — | |
| FTModel=Gemma-2-9B2026.01 | 0.9333 | — | |
| IDGAFModel=Llama-3.2-1B2026.01 | 0.9303 | — | |
| Refusal SuppressionModel=Gemma-2-9B2026.01 | 0.9273 | — | |
| FormalLLM Model=Qwen2.5-32B2026.01 | 0.9273 | — | |
| FTModel=Gemma-2-2B2026.01 | 0.9273 | — | |
| FormalLLM Model=Llama-3.3-70B2026.01 | 0.9242 | — | |
| BaseModel=Qwen3-4B-think2026.01 | 0.9212 | — | |
| BaseLLM Model=Llama-3.3-70B2026.01 | 0.9212 | — |