Response Harmfulness Detection on SafeRLHF
72.1F1 ScoreBeaverDam
Evaluation Results
| Method | Links | |
|---|---|---|
| BeaverDamModel Size=7B2026.05 | 72.1 | |
| COLAGUARDModel Size=8B2026.05 | 70.49 | |
| GuardReasonerModel Size=8B2026.05 | 70.04 | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=8B2026.05 | 70.04 | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=7B2026.05 | 69.98 | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 69.7 | |
| GuardReasonerModel Size=3B2026.05 | 69.02 | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 69.02 | |
| COLAGUARDModel Size=3B2026.05 | 68.72 | |
| GuardReasonerModel Size=1B2026.05 | 68.39 | |
| Claude 3.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 67.54 | |
| o1-previewModel Size=Unknown2026.05 | 66.6 | |
| GPT-4o+CoTModel Size=Unknown2026.05 | 65.1 | |
| o1-pre+CoTCategory=CoT-Based LRM, Model Size=-2026.05 | 65.1 | |
| MD-JudgeModel Size=7B2026.05 | 64.7 | |
| MD-JudgeCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 64.7 | |
| qwen3-CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 64.42 | |
| GPT-4o+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 64.34 | |
| WildGuardModel Size=7B2026.05 | 64.2 | |
| WildGuardCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 64.2 | |
| GPT-4oModel Size=Unknown2026.05 | 64.05 | |
| Gemini 1.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 63.73 | |
| QWQ+CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 63.71 | |
| QwQ-previewModel Size=32B2026.05 | 62.76 | |
| GPT-4+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 61.25 | |
| HarmBench LLaMAModel Size=13B2026.05 | 60 | |
| Aegis Guard DefensiveModel Size=7B2026.05 | 59.3 | |
| Aegis Guard DefCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 59.3 | |
| Aegis Guard PermissiveModel Size=7B2026.05 | 55.9 | |
| Aegis Guard PerCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 55.9 | |
| HarmBench MistralModel Size=7B2026.05 | 52.4 | |
| LLaMA Guard 2Model Size=8B2026.05 | 51.6 | |
| LLaMA Guard 2Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 51.6 | |
| LLaMA GuardModel Size=7B2026.05 | 48.4 | |
| ShieldGemmaModel Size=9B2026.05 | 47.07 | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=9B2026.05 | 47.07 | |
| LLaMA Guard 3Model Size=8B2026.05 | 44.36 | |
| LLaMA Guard 3Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 44.36 | |
| ShieldGemmaModel Size=2B2026.05 | 16.92 | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=2B2026.05 | 16.92 | |
| ModerationCategory=Discriminant-based Guardrails, Model Size=-2026.05 | 10.1 |