Jailbreak Detection on GoalFrameBench
94AccuracyFrameShield-Crit
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FrameShield-CritLLM Backbone=Qwen2.5-7B2026.02 | 94 | 93 | |
| FrameShield-CritLLM Backbone=Qwen3-4B2026.02 | 89 | 87 | |
| FrameShield-LastLLM Backbone=Llama3-8B2026.02 | 82 | 79 | |
| FrameShield-CritLLM Backbone=Vicuna-7B2026.02 | 81 | 79 | |
| FrameShield-CritLLM Backbone=Vicuna-13B2026.02 | 81 | 82 | |
| FrameShield-CritLLM Backbone=Llama2-7B2026.02 | 80 | 79 | |
| FrameShield-LastLLM Backbone=Mistral-7B2026.02 | 79 | 74 | |
| FrameShield-LastLLM Backbone=Qwen2.5-7B2026.02 | 79 | 76 | |
| FrameShield-LastLLM Backbone=Vicuna-13B2026.02 | 76 | 79 | |
| FrameShield-LastLLM Backbone=Qwen3-4B2026.02 | 76 | 75 | |
| FrameShield-CritLLM Backbone=Llama3-8B2026.02 | 76 | 69 | |
| FrameShield-CritLLM Backbone=Mistral-7B2026.02 | 76 | 69 | |
| JBShieldLLM Backbone=Llama3-8B2026.02 | 75 | 75 | |
| FrameShield-CritLLM Backbone=Qwen2-0.5B2026.02 | 74 | 65 | |
| JBShieldLLM Backbone=Llama2-7B2026.02 | 73 | 76 | |
| JBShieldLLM Backbone=Mistral-7B2026.02 | 71 | 75 | |
| JBShieldLLM Backbone=Qwen3-4B2026.02 | 70 | 74 | |
| FrameShield-LastLLM Backbone=Vicuna-7B2026.02 | 67 | 75 | |
| JBShieldLLM Backbone=Qwen2.5-7B2026.02 | 66 | 68 | |
| FrameShield-LastLLM Backbone=Qwen2-0.5B2026.02 | 66 | 74 | |
| JBShieldLLM Backbone=Vicuna-13B2026.02 | 63 | 67 | |
| JBShieldLLM Backbone=Vicuna-7B2026.02 | 60 | 63 | |
| JBShieldLLM Backbone=Qwen2-0.5B2026.02 | 59 | 62 | |
| FrameShield-LastLLM Backbone=Llama2-7B2026.02 | 58 | 70 |