Safety Classification on ToxicChat
0.81F1 ScoreQwen3Guard
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen3GuardSize=8B, Mode=loose, Reasoning=false2025.12 | 0.81 | |
| Label and intent rewardOptimization Strategy=GRPO2026.06 | 0.743 | |
| Nemotron Safety 4BOptimization Strategy=Dedicated Safety Guard2026.06 | 0.733 | |
| LE-DPOOptimization Strategy=DPO2026.06 | 0.733 | |
| AprielGuardSize=8B, Reasoning=false2025.12 | 0.73 | |
| Gemma-3-12BOptimization Strategy=SFT, Supervision Source=Annotated Intents2026.06 | 0.727 | |
| AprielGuardSize=8B, Reasoning=true2025.12 | 0.72 | |
| IBM Granite Guardian 3.3Size=8B, Reasoning=false2025.12 | 0.71 | |
| IF-DPOOptimization Strategy=DPO2026.06 | 0.708 | |
| ShieldGemma 27BOptimization Strategy=Dedicated Safety Guard2026.06 | 0.703 | |
| Human-intent on AIMSOptimization Strategy=Reasoning Distillation, Teacher Model=GPT-OSS-120B, Student Model=Gemma-3-12B2026.06 | 0.702 | |
| gpt-oss-safeguardSize=20B, Reasoning=true2025.12 | 0.7 | |
| Label rewardOptimization Strategy=GRPO2026.06 | 0.685 | |
| GuardReasoner 8BOptimization Strategy=Dedicated Safety Guard2026.06 | 0.681 | |
| IBM Granite Guardian 3.2Size=5B, Reasoning=false2025.12 | 0.68 | |
| GPT-5.4Optimization Strategy=Zero-shot2026.06 | 0.676 | |
| ShieldGemmaSize=9B, Reasoning=false2025.12 | 0.67 | |
| Claude Sonnet 4.6Optimization Strategy=Zero-shot2026.06 | 0.667 | |
| Llama-3.1-8BOptimization Strategy=SFT, Supervision Source=Annotated Intents2026.06 | 0.664 | |
| WildGuard 7BOptimization Strategy=Dedicated Safety Guard2026.06 | 0.652 | |
| Gemma-3-12BOptimization Strategy=Zero-shot2026.06 | 0.644 | |
| GPT-OSS-Safeguard 120BOptimization Strategy=Dedicated Safety Guard2026.06 | 0.643 | |
| GPT-OSS-120BOptimization Strategy=Zero-shot2026.06 | 0.641 | |
| Qwen3GuardSize=8B, Mode=strict, Reasoning=false2025.12 | 0.63 | |
| IBM Granite Guardian 3.3Size=8B, Reasoning=true2025.12 | 0.63 | |
| Llama-3.1-8BOptimization Strategy=Zero-shot2026.06 | 0.516 | |
| IBM Granite Guardian 3.1Size=2B, Reasoning=false2025.12 | 0.51 | |
| Llama Guard 3Reasoning=false2025.12 | 0.48 | |
| IBM Granite Guardian 3.2Size=3B, Reasoning=false2025.12 | 0.45 | |
| LlamaGuard 4Optimization Strategy=Dedicated Safety Guard2026.06 | 0.441 | |
| Llama Guard 4Reasoning=false2025.12 | 0.43 | |
| Llama Guard 2Reasoning=false2025.12 | 0.42 |