Safety Classification on XSTest
95.8F1 ScoreLabel and intent reward
Evaluation Results
| Method | Links | |
|---|---|---|
| Label and intent rewardOptimization Strategy=GRPO2026.06 | 95.8 | |
| YuFeng-XGuard-Reason-8BParams=8B2026.07 | 95.4 | |
| WildGuard 7BOptimization Strategy=Dedicated Safety Guard2026.06 | 94.5 | |
| GPT-OSS-Safeguard 120BOptimization Strategy=Dedicated Safety Guard2026.06 | 94.4 | |
| AprielGuardSize=8B, Reasoning=false2025.12 | 94 | |
| Human-intent on AIMSOptimization Strategy=Reasoning Distillation, Teacher Model=GPT-OSS-120B, Student Model=Gemma-3-12B2026.06 | 93.6 | |
| Stage1-SFT-v4Params=4B2026.07 | 93.6 | |
| Stage3-DPOParams=4B2026.07 | 92.9 | |
| GPT-5.4Optimization Strategy=Zero-shot2026.06 | 92 | |
| GuardReasoner 8BOptimization Strategy=Dedicated Safety Guard2026.06 | 91.9 | |
| YuFeng-XGuard-Reason-0.6BParams=0.6B2026.07 | 91.8 | |
| GPT-OSS-120BOptimization Strategy=Zero-shot2026.06 | 91.1 | |
| Qwen3GuardSize=8B, Mode=strict, Reasoning=false2025.12 | 91 | |
| AprielGuardSize=8B, Reasoning=true2025.12 | 91 | |
| IF-DPOOptimization Strategy=DPO2026.06 | 90.9 | |
| Llama-3.1-8BOptimization Strategy=SFT, Supervision Source=Annotated Intents2026.06 | 90.8 | |
| Qwen3Guard-8B-GenParams=8B2026.07 | 90.8 | |
| Llama-3.1-8BOptimization Strategy=Zero-shot2026.06 | 90.4 | |
| Label rewardOptimization Strategy=GRPO2026.06 | 90.4 | |
| Gemma-3-12BOptimization Strategy=Zero-shot2026.06 | 90.2 | |
| Qwen3GuardSize=8B, Mode=loose, Reasoning=false2025.12 | 90 | |
| gpt-oss-safeguardSize=20B, Reasoning=true2025.12 | 90 | |
| Qwen3Guard-4B-GenParams=4B2026.07 | 89.9 | |
| Llama Guard 2Reasoning=false2025.12 | 89 | |
| Gemma-3-12BOptimization Strategy=SFT, Supervision Source=Annotated Intents2026.06 | 88.4 | |
| LE-DPOOptimization Strategy=DPO2026.06 | 88.4 | |
| Llama Guard 3Reasoning=false2025.12 | 88 | |
| Stage1-SFT-v1Params=4B2026.07 | 87.4 | |
| Stage1-SFT-v0Params=4B2026.07 | 87.1 | |
| IBM Granite Guardian 3.3Size=8B, Reasoning=true2025.12 | 87 | |
| Stage2-SFTParams=4B2026.07 | 86.7 | |
| IBM Granite Guardian 3.3Size=8B, Reasoning=false2025.12 | 86 | |
| Claude Sonnet 4.6Optimization Strategy=Zero-shot2026.06 | 86 | |
| Stage1-SFT-v3Params=4B2026.07 | 86 | |
| Qwen3Guard-0.8B-GenParams=0.8B2026.07 | 85.3 | |
| Nemotron Safety 4BOptimization Strategy=Dedicated Safety Guard2026.06 | 85.1 | |
| IBM Granite Guardian 3.2Size=5B, Reasoning=false2025.12 | 85 | |
| Stage1-SFT-v2Params=4B2026.07 | 84.7 | |
| Llama Guard 4Reasoning=false2025.12 | 84 | |
| LlamaGuard 4Optimization Strategy=Dedicated Safety Guard2026.06 | 83.6 | |
| IBM Granite Guardian 3.1Size=2B, Reasoning=false2025.12 | 83 | |
| ShieldGemma 27BOptimization Strategy=Dedicated Safety Guard2026.06 | 82.3 | |
| ShieldGemmaSize=9B, Reasoning=false2025.12 | 82 | |
| IBM Granite Guardian 3.2Size=3B, Reasoning=false2025.12 | 81 | |
| CREST-LARGEModel Variant=Large, Backbone=XLM-RoBERTa-Large2025.12 | 69.83 | |
| CREST-BASEModel Variant=Base, Backbone=XLM-RoBERTa-Base2025.12 | 67.04 |