List your company
ResearchFormerly WizWand

Harmful prompt detection on OpenAI

81.3F1 ScoreStage1-SFT-v2
63.921668.433372.94577.4567Feb 22, 2025May 16, 2025Aug 7, 2025Oct 30, 2025Jan 21, 2026Apr 14, 2026Jul 7, 2026
Updated 3mo ago

Evaluation results

MethodLinks
Stage1-SFT-v2Params=4B2026.0781.3
Stage1-SFT-v4Params=4B2026.0780.3
Stage2-SFTParams=4B2026.0779.8
LlamaGuard3Methodology=Guard Model2025.0279.11
Stage3-DPOParams=4B2026.0779
GraniteGuardian-3-1-8BMethodology=Guard Model2025.0277.63
ShieldGemma-9BMethodology=Guard Model2025.0277.63
Stage1-SFT-v1Params=4B2026.0776.8
Aegis-Guard-DMethodology=Guard Model2025.0276.44
Stage1-SFT-v3Params=4B2026.0775.8
Ayub & MajumdarBackbone=Mistral-7B-Inst, Methodology=Latent-Based2025.0275.69
Abdelnabi et al.Backbone=OLMo2-7B-Inst, Methodology=Latent-Based2025.0275.19
YuFeng-XGuard-Reason-8BParams=8B2026.0774.7
Ayub & MajumdarBackbone=Qwen3-8B-Inst, Methodology=Latent-Based2025.0274.56
MLPMBackbone=OLMo2-7B-Inst, Methodology=Latent-Based2025.0274.21
Stage1-SFT-v0Params=4B2026.0774
MLPMBackbone=Llama-8B-Inst, Methodology=Latent-Based2025.0272.85
YuFeng-XGuard-Reason-0.6BParams=0.6B2026.0772.8
MLPMBackbone=Qwen3-8B-Inst, Methodology=Latent-Based2025.0272.35
WildGuardMethodology=Guard Model2025.0272.28
MLPMBackbone=Mistral-7B-Inst, Methodology=Latent-Based2025.0270.68
Qwen3Guard-8B-GenParams=8B2026.0768.5
Abdelnabi et al.Backbone=Qwen3-8B-Inst, Methodology=Latent-Based2025.0268.45
Qwen3Guard-4B-GenParams=4B2026.0768.4
Abdelnabi et al.Backbone=Llama-8B-Inst, Methodology=Latent-Based2025.0267.99
Ayub & MajumdarBackbone=OLMo2-7B-Inst, Methodology=Latent-Based2025.0266.95
Ayub & MajumdarBackbone=Llama-8B-Inst, Methodology=Latent-Based2025.0266.6
Qwen3Guard-0.8B-GenParams=0.8B2026.0766.2
Abdelnabi et al.Backbone=Mistral-7B-Inst, Methodology=Latent-Based2025.0264.59