Response Harmfulness Detection on HarmBench
98.94F1 ScoreLLaMA Guard 3
Evaluation Results
| Method | Links | |
|---|---|---|
| LLaMA Guard 3Model Size=8B2026.05 | 98.94 | |
| LLaMA Guard 3Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 98.94 | |
| WildGuardModel Size=7B2026.05 | 98.9 | |
| WildGuardCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 98.72 | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 98.29 | |
| GuardReasonerModel Size=1B2026.05 | 96.31 | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=7B2026.05 | 95.79 | |
| COLAGUARDModel Size=3B2026.05 | 94.25 | |
| LLaMA Guard 2Model Size=8B2026.05 | 94 | |
| LLaMA Guard 2Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 94 | |
| COLAGUARDModel Size=8B2026.05 | 93.54 | |
| GuardReasonerModel Size=8B2026.05 | 91.86 | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=8B2026.05 | 91.86 | |
| o1-previewModel Size=Unknown2026.05 | 89.61 | |
| GuardReasonerModel Size=3B2026.05 | 89.1 | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 89.1 | |
| GuardReasoner-Omni 4BModel Category=VLM Guard Model2026.02 | 87.61 | |
| o1-pre+CoTCategory=CoT-Based LRM, Model Size=-2026.05 | 87.14 | |
| HarmBench Mistral 7BModel Category=LLM Guard Model2026.02 | 87 | |
| HarmBench MistralModel Size=7B2026.05 | 87 | |
| Qwen3Guard Gen 8BModel Category=LLM Guard Model, Detection Mode=strict2026.02 | 86.91 | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 86.88 | |
| GuardReasoner-VL 7BModel Category=VLM Guard Model2026.02 | 86.82 | |
| QwQ-previewModel Size=32B2026.05 | 86.73 | |
| Qwen3Guard Gen 8BModel Category=LLM Guard Model, Detection Mode=loose2026.02 | 86.68 | |
| GuardReasoner-Omni 2BModel Category=VLM Guard Model2026.02 | 86.46 | |
| COLAGUARDModel Size=8B2026.05 | 86.38 | |
| COLAGUARDModel Size=3B2026.05 | 86.36 | |
| WildGuard 7BModel Category=LLM Guard Model2026.02 | 86.3 | |
| WildGuardModel Size=7B2026.05 | 86.3 | |
| GuardReasoner-VL 3BModel Category=VLM Guard Model2026.02 | 86.21 | |
| WildGuardCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 85.97 | |
| GuardReasonerModel Size=3B2026.05 | 85.66 | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 85.66 | |
| QWQ+CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 85.61 | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=7B2026.05 | 85.6 | |
| GuardReasoner 8BModel Category=LLM Guard Model2026.02 | 85.47 | |
| GuardReasonerModel Size=8B2026.05 | 85.47 | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=8B2026.05 | 85.47 | |
| LLaMA Guard 3 8BModel Category=LLM Guard Model2026.02 | 85.07 | |
| LLaMA Guard 3Model Size=8B2026.05 | 85.07 | |
| LLaMA Guard 3Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 85.07 | |
| qwen3-CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 85.01 | |
| DistilRoBERTa-based NBFModel Size=87M2025.02 | 84.8 | |
| GuardReasonerModel Size=1B2026.05 | 84.75 | |
| HarmBench LLaMA 13BModel Category=LLM Guard Model2026.02 | 84.3 | |
| HarmBench LLaMAModel Size=13B2026.05 | 84.3 | |
| GPT-4+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 84 | |
| Gemini 1.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 83.35 | |
| LLaMA Guard 4 12BModel Category=VLM Guard Model2026.02 | 82.92 | |
| QWQ+CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 82.64 | |
| GPT-4oModel Size=Unknown2026.05 | 82.27 | |
| GPT-4o+CoTModel Size=Unknown2026.05 | 81.98 | |
| MD-Judge 7BModel Category=LLM Guard Model2026.02 | 81.6 | |
| MD-JudgeModel Size=7B2026.05 | 81.6 | |
| MD-JudgeCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 81.6 | |
| GPT-4o+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 81.46 | |
| MPNet-based NBFModel Size=115M2025.02 | 81.1 | |
| Claude 3.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 80.6 | |
| qwen3-CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 80.1 | |
| Gemini 1.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 80.06 | |
| GPT-4+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 78.83 | |
| LLaMA Guard 2Model Size=8B2026.05 | 77.8 | |
| LLaMA Guard 2Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 77.8 | |
| Aegis Guard DefensiveModel Size=7B2026.05 | 77.7 | |
| Aegis Guard DefCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 77.7 | |
| Claude 3.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 77.29 | |
| o1-previewModel Size=Unknown2026.05 | 76.4 | |
| o1-pre+CoTCategory=CoT-Based LRM, Model Size=-2026.05 | 75.77 | |
| PolyGuard-Qwen 7BModel Category=LLM Guard Model2026.02 | 71.14 | |
| Aegis Guard PermissiveModel Size=7B2026.05 | 70.5 | |
| Aegis Guard PerCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 70.5 | |
| QwQ-previewModel Size=32B2026.05 | 69.65 | |
| GPT-4o+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 68.7 | |
| ShieldGemmaModel Size=9B2026.05 | 67.96 | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=9B2026.05 | 67.96 | |
| LLaMA GuardModel Size=7B2025.02 | 67.2 | |
| LLaMA GuardModel Size=7B2026.05 | 67.2 | |
| GPT-4o+CoTModel Size=Unknown2026.05 | 65.99 | |
| Aegis Guard Defensive 7BModel Category=LLM Guard Model2026.02 | 62.2 | |
| Aegis Guard DefensiveModel Size=7B2026.05 | 62.2 | |
| Aegis Guard DefCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 62.2 | |
| Aegis Guard Permissive 7BModel Category=LLM Guard Model2026.02 | 60.8 | |
| Aegis Guard PermissiveModel Size=7B2026.05 | 60.8 | |
| Aegis Guard PerCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 60.8 | |
| BeaverDam 7BModel Category=LLM Guard Model2026.02 | 58.4 | |
| BeaverDamModel Size=7B2026.05 | 58.4 | |
| ShieldGemma 9BModel Category=LLM Guard Model2026.02 | 56.44 | |
| ShieldGemmaModel Size=9B2026.05 | 56.44 | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=9B2026.05 | 56.44 | |
| GPT-4oModel Size=Unknown2026.05 | 56.34 | |
| LLaMA GuardModel Size=7B2026.05 | 52 | |
| ShieldGemmaModel Size=2B2026.05 | 35.36 | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=2B2026.05 | 35.36 | |
| ModerationCategory=Discriminant-based Guardrails, Model Size=-2026.05 | 20.6 | |
| ShieldGemmaModel Size=2B2026.05 | 11.81 | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=2B2026.05 | 11.81 | |
| ShieldGemmaModel Size=2B2025.02 | 11.8 | |
| OpenAI ModerationModel Size=Unknown2025.02 | 9.6 | |
| ModerationCategory=Discriminant-based Guardrails, Model Size=-2026.05 | 9.6 |