Prompt classification on Aegis
91.5F1 ScoreStage3-DPO
Evaluation Results
| Method | Links | |
|---|---|---|
| Stage3-DPOParams=4B2026.07 | 91.5 | |
| Qwen3Guard-GenSize=8B2026.07 | 91.4 | |
| Stage1-SFT-v0Params=4B2026.07 | 91.3 | |
| Qwen3Guard-8B-GenParams=8B2026.07 | 91.2 | |
| Stage1-SFT-v1Params=4B2026.07 | 91 | |
| Qwen3Guard-GenSize=0.6B2026.07 | 90.8 | |
| Qwen3Guard-GenSize=4B2026.07 | 90.8 | |
| Qwen3Guard-0.8B-GenParams=0.8B2026.07 | 90.6 | |
| Qwen3Guard-4B-GenParams=4B2026.07 | 90.6 | |
| PolyGuard-QwenSize=7B2026.07 | 90.3 | |
| Stage1-SFT-v3Params=4B2026.07 | 90.2 | |
| NemotronReasoningModel Size=4B2026.01 | 89.6 | |
| PolyGuardModel Size=7B2026.01 | 89.6 | |
| YuFeng-XGuardModel Size=8B2026.01 | 89.6 | |
| WildGuardSize=7B2026.07 | 89.4 | |
| NExT-GuardSafeguard Type=Streaming2026.02 | 88.9 | |
| Qwen3Guard-GenModel Size=8B, Evaluation protocol=strict2026.01 | 88.7 | |
| NExT-Guard*Safeguard Type=Streaming2026.02 | 88.5 | |
| YuFeng-XGuardModel Size=0.6B2026.01 | 88.3 | |
| HaloGuard 1.0Size=4B2026.07 | 88 | |
| Qwen3Guard-GenModel Size=0.6B, Evaluation protocol=strict2026.01 | 87.9 | |
| WildGuard-7BSafeguard Type=Post-hoc2026.02 | 87.9 | |
| WildGuardModel Size=7B2026.01 | 87.6 | |
| Stage2-SFTParams=4B2026.07 | 87.6 | |
| Stage1-SFT-v2Params=4B2026.07 | 87.5 | |
| Qwen3Guard-GenModel Size=4B, Evaluation protocol=strict2026.01 | 87.4 | |
| YuFeng-XGuard-Reason-0.6BParams=0.6B2026.07 | 87 | |
| YuFeng-XGuard-Reason-8BParams=8B2026.07 | 86.9 | |
| HaloGuard 1.0Size=0.8B2026.07 | 86.7 | |
| Stage1-SFT-v4Params=4B2026.07 | 85 | |
| GPT-OSS-SafeGuardModel Size=20B2026.01 | 84.3 | |
| NemotronGuardV2Model Size=8B2026.01 | 82.2 | |
| Qwen3Guard-GenModel Size=0.6B, Evaluation protocol=loose2026.01 | 82.1 | |
| Qwen3Guard-GenModel Size=8B, Evaluation protocol=loose2026.01 | 81.8 | |
| NemoGuardSize=8B2026.07 | 81.4 | |
| Qwen3Guard-GenModel Size=4B, Evaluation protocol=loose2026.01 | 81.2 | |
| NemoGuard-8BSafeguard Type=Post-hoc2026.02 | 80.8 | |
| ShieldGemmaModel Size=9B2026.01 | 79.8 | |
| Llama3GuardModel Size=8B2026.01 | 77.8 | |
| Qwen3Guard-0.6B-StreamSafeguard Type=Streaming2026.02 | 76.2 | |
| Qwen3Guard-0.6B-GenSafeguard Type=Post-hoc2026.02 | 75.5 | |
| Qwen3Guard-8B-GenSafeguard Type=Post-hoc2026.02 | 75 | |
| Qwen3Guard-4B-GenSafeguard Type=Post-hoc2026.02 | 74.9 | |
| Qwen3Guard-8B-StreamSafeguard Type=Streaming2026.02 | 74.9 | |
| Qwen3Guard-4B-StreamSafeguard Type=Streaming2026.02 | 74.3 | |
| SCM-7BSafeguard Type=Streaming2026.02 | 73.8 | |
| KelpSafeguard Type=Streaming2026.02 | 73.5 | |
| Llama4GuardModel Size=12B2026.01 | 72.2 | |
| LlamaGuard3-8BSafeguard Type=Post-hoc2026.02 | 70.3 | |
| ShieldGemmaSize=27B2026.07 | 69 | |
| ShieldGemma-9BSafeguard Type=Post-hoc2026.02 | 68.9 | |
| ShieldGemma-27BSafeguard Type=Post-hoc2026.02 | 68.1 | |
| LlamaGuard4Size=12B2026.07 | 67.8 | |
| LlamaGuard4-12BSafeguard Type=Post-hoc2026.02 | 66.5 |