Toxicity Detection on ToxicChat
1F1 ScoreSafeGPT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SafeGPT2026.01 | 1 | 1 | 1 | 0 | — | |
| ReflectiveProposerUnderlying LM=Qwen3-14B, Feedback=REFLECTIVEPROPOSER2026.01 | 0.907 | 0.916 | 0.905 | — | — | |
| UnoptimizedUnderlying LM=Qwen3-14B, Feedback=None2026.01 | 0.849 | 0.869 | 0.844 | — | — | |
| MIPROUnderlying LM=Qwen3-14B, Feedback=PREDICTOR & Interp.2026.01 | 0.796 | 0.844 | 0.791 | — | — | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=7B2026.05 | 0.7905 | — | — | — | — | |
| GuardReasonerModel Size=8B2026.05 | 0.7879 | — | — | — | — | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=8B2026.05 | 0.7879 | — | — | — | — | |
| GuardReasonerModel Size=3B2026.05 | 0.782 | — | — | — | — | |
| GuardReasonerCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 0.782 | — | — | — | — | |
| ConsisGuardCategory=Reasoning-based Guardrails, Model Size=3B2026.05 | 0.7726 | — | — | — | — | |
| COLAGUARDModel Size=3B2026.05 | 0.7527 | — | — | — | — | |
| COLAGUARDModel Size=8B2026.05 | 0.7526 | — | — | — | — | |
| MIPROUnderlying LM=Qwen3-14B, Feedback=PREDICTOR2026.01 | 0.75 | 0.819 | 0.749 | — | — | |
| GPT-4o+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 0.7452 | — | — | — | — | |
| GPT-4o+CoTModel Size=Unknown2026.05 | 0.7343 | — | — | — | — | |
| Aegis Guard PermissiveModel Size=7B2026.05 | 0.73 | — | — | — | — | |
| Aegis Guard PerCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 0.73 | — | — | — | — | |
| GuardReasonerModel Size=1B2026.05 | 0.7243 | — | — | — | — | |
| WildGuardModel Size=7B2026.05 | 0.708 | — | — | — | — | |
| WildGuardCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 0.708 | — | — | — | — | |
| Aegis Guard DefensiveModel Size=7B2026.05 | 0.7 | — | — | — | — | |
| Aegis Guard DefCategory=Discriminant-based Guardrails, Model Size=7B2026.05 | 0.7 | — | — | — | — | |
| qwen3-CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 0.6962 | — | — | — | — | |
| ShieldGemmaModel Size=9B2026.05 | 0.6792 | — | — | — | — | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=9B2026.05 | 0.6792 | — | — | — | — | |
| Gemini 1.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 0.666 | — | — | — | — | |
| QWQ+CoTCategory=CoT-Based LRM, Model Size=32B2026.05 | 0.6653 | — | — | — | — | |
| GPT-4oModel Size=Unknown2026.05 | 0.6446 | — | — | — | — | |
| GPT-4+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 0.6396 | — | — | — | — | |
| o1-pre+CoTCategory=CoT-Based LRM, Model Size=-2026.05 | 0.637 | — | — | — | — | |
| LLaMA GuardModel Size=7B2026.05 | 0.616 | — | — | — | — | |
| o1-previewModel Size=Unknown2026.05 | 0.5769 | — | — | — | — | |
| Claude 3.5+CoTCategory=CoT-Based LLM, Model Size=-2026.05 | 0.5458 | — | — | — | — | |
| LLaMA Guard 3Model Size=8B2026.05 | 0.5312 | — | — | — | — | |
| LLaMA Guard 3Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 0.5312 | — | — | — | — | |
| LLaMA Guard 2Model Size=8B2026.05 | 0.471 | — | — | — | — | |
| LLaMA Guard 2Category=Discriminant-based Guardrails, Model Size=8B2026.05 | 0.471 | — | — | — | — | |
| QwQ-previewModel Size=32B2026.05 | 0.3481 | — | — | — | — | |
| ModerationCategory=Discriminant-based Guardrails, Model Size=-2026.05 | 0.254 | — | — | — | — | |
| ShieldGemmaModel Size=2B2026.05 | 0.0691 | — | — | — | — | |
| ShieldGemmaCategory=Discriminant-based Guardrails, Model Size=2B2026.05 | 0.0691 | — | — | — | — | |
| Regex-Only2026.01 | 0 | 0 | 0 | 0 | — | |
| Simple NER2026.01 | 0 | 0 | 0 | 0 | — | |
| Keyword2026.01 | 0 | 0 | 0 | 0 | — | |
| Hybrid2026.01 | 0 | 0 | 0 | 0 | — | |
| LlamaGuard3Backbone=Llama3.2-1B2026.04 | — | — | — | — | 0.633 | |
| LlamaGuard3Backbone=Llama3.1-8B2026.04 | — | — | — | — | 0.722 | |
| Qwen3GuardBackbone=Qwen3-0.6B2026.04 | — | — | — | — | 0.82 | |
| Qwen3GuardBackbone=Qwen3-4B2026.04 | — | — | — | — | 0.849 | |
| SIRENBackbone=Qwen3-0.6B2026.04 | — | — | — | — | 0.816 | |
| SIRENBackbone=Llama3.2-1B2026.04 | — | — | — | — | 0.8 | |
| SIRENBackbone=Qwen3-4B2026.04 | — | — | — | — | 0.835 | |
| SIRENBackbone=Llama3.1-8B2026.04 | — | — | — | — | 0.831 |