List your company
ResearchFormerly WizWand

Safety Guardrailing on AlpacaEval

0False Positive RateLlama Guard
-3.9722.827549.62576.4225Apr 1, 2026
Updated 7mo ago

Evaluation results

MethodLinks
Llama GuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Post-checking2026.040
GradientCuffBackbone=Qwen3.5-9B, Average=0.002026.040
SelfGraderBackbone=Qwen3.5-9B, Average=0.002026.040
Llama Guard (Post)Backbone=Qwen3.5-9B, Average=0.432026.040.1
Llama GuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Pre-checking2026.040.37
Llama Guard (Pre)Backbone=Qwen3.5-9B, Average=1.452026.040.4
WildGuard (Post)Backbone=Qwen3.5-9B, Average=0.182026.040.4
GradSafeBase Model=Llama-3-8B-Instruct2026.040.49
WildGuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Post-checking2026.040.49
Perplexity FilterBackbone=Qwen3.5-9B, Average=0.152026.040.6
Perplexity FilterBase Model=Llama-3-8B-Instruct2026.040.62
Prompt GuardBase Model=Llama-3-8B-Instruct2026.040.62
SelfDefendBase Model=Llama-3-8B-Instruct, Defense Mode=Intent2026.040.74
Prompt GuardBackbone=Qwen3.5-9B, Average=0.982026.040.8
SelfDefend (Intent)Backbone=Qwen3.5-9B, Average=2.332026.040.9
GuardReasonerBase Model=Llama-3-8B-Instruct, Guardrail Stage=Post-checking2026.040.99
SelfGraderBase Model=Llama-3-8B-Instruct2026.041.36
GuardReasonerBase Model=Llama-3-8B-Instruct, Guardrail Stage=Pre-checking2026.041.49
WildGuard (Pre)Backbone=Qwen3.5-9B, Average=3.602026.042.7
WildGuardBase Model=Llama-3-8B-Instruct, Guardrail Stage=Pre-checking2026.042.73
SelfDefendBase Model=Llama-3-8B-Instruct, Defense Mode=Direct2026.043.72
SelfDefend (Direct)Backbone=Qwen3.5-9B, Average=7.332026.044.5
GradientCuffBase Model=Llama-3-8B-Instruct2026.0424.09
Token HighlighterBase Model=Llama-3-8B-Instruct2026.0499.25