Safety Classification on WildGuard (test)
88.8F1 ScoreWildGuard 7B
Evaluation Results
| Method | Links | |
|---|---|---|
| WildGuard 7BOptimization Strategy=Dedicated Safety Guard2026.06 | 88.8 | |
| GuardReasoner 8BOptimization Strategy=Dedicated Safety Guard2026.06 | 88.8 | |
| TPCParams=1,381,8892025.09 | 88.5 | |
| GPT-OSS-120BOptimization Strategy=Zero-shot2026.06 | 88.4 | |
| MLP probeParams=1,382,1472025.09 | 88.2 | |
| GPT-5.4Optimization Strategy=Zero-shot2026.06 | 88 | |
| Bilinear probeParams=344,1292025.09 | 87.82 | |
| EE-MLPParams=1,381,9182025.09 | 87.76 | |
| Human-intent on AIMSOptimization Strategy=Reasoning Distillation, Teacher Model=GPT-OSS-120B, Student Model=Gemma-3-12B2026.06 | 87.6 | |
| GPT-OSS-Safeguard 120BOptimization Strategy=Dedicated Safety Guard2026.06 | 87.1 | |
| Label rewardOptimization Strategy=GRPO2026.06 | 87.1 | |
| Linear probeParams=5,3772025.09 | 86.86 | |
| gpt-4o-miniParams=≈ 8,000,000,000*2025.09 | 86.63 | |
| Label and intent rewardOptimization Strategy=GRPO2026.06 | 86.3 | |
| Gemma-3-12BOptimization Strategy=SFT, Supervision Source=Annotated Intents2026.06 | 85.7 | |
| Llama-3.1-8BOptimization Strategy=SFT, Supervision Source=Annotated Intents2026.06 | 85.6 | |
| LE-DPOOptimization Strategy=DPO2026.06 | 85.6 | |
| Gemma-3-12BOptimization Strategy=Zero-shot2026.06 | 85.3 | |
| Nemotron Safety 4BOptimization Strategy=Dedicated Safety Guard2026.06 | 85.2 | |
| IF-DPOOptimization Strategy=DPO2026.06 | 85.1 | |
| Claude Sonnet 4.6Optimization Strategy=Zero-shot2026.06 | 83.8 | |
| claude-3-haikuParams=≈ 8,000,000,000*2025.09 | 83.24 | |
| o3-miniParams=unknown2025.09 | 82.03 | |
| SentGuardbackbone=Qwen3-4B-Instruct-25072026.06 | 80 | |
| Qwen3Guard-8B2026.06 | 77.2 | |
| Llama-Guard-3-8BParams=8,030,261,2482025.09 | 76.85 | |
| Llama-3.1-8BOptimization Strategy=Zero-shot2026.06 | 76.2 | |
| WildGuard-7B2026.06 | 75.3 | |
| GPT-5.5protocol=zero-shot2026.06 | 74.4 | |
| LlamaGuard 4Optimization Strategy=Dedicated Safety Guard2026.06 | 73.8 | |
| LlamaGuard3-8B2026.06 | 70.7 | |
| shieldgemma-2bThreshold=optimal, Params=2,614,341,8882025.09 | 69.9 | |
| LlamaGuard4-12B2026.06 | 66.7 | |
| Gemini-3.5-Flashprotocol=zero-shot2026.06 | 59.4 | |
| ShieldGemma 27BOptimization Strategy=Dedicated Safety Guard2026.06 | 51.2 |