Security Interception on Bilingual Security Benchmark Chinese
98.2Interception Rate (IR)Qwen2.5-14B
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Qwen2.5-14BCondition=Qwen2.5-14B2026.03 | 98.2 | 9.7 | 94.1 | 100 | 98.3 | 96.3 | 77.5 | |
| Llama-3.1-8BCondition=Llama-3.1-8B2026.03 | 98 | 37.5 | 76.4 | 99.3 | 96.7 | 98.5 | 99.2 | |
| Qwen3.5-9BCondition=Qwen3.5-9B2026.03 | 96.5 | 34.1 | 78.3 | 97.2 | 94.4 | 98.5 | 77.5 | |
| Qwen3.5-4BCondition=Qwen3.5-4B2026.03 | 95.8 | 27.5 | 82.5 | 97.9 | 96.1 | 93.4 | 86.7 | |
| Cascade (Q→14B)Condition=Cascade (Q→14B)2026.03 | 95.6 | 6 | 94.8 | 97.2 | 93.9 | 96.3 | — | |
| GPT-4o-miniCondition=GPT-4o-mini2026.03 | 93 | 3.2 | 94.9 | 92.9 | 92.8 | 93.4 | 75 | |
| Cascade (Q→G)Condition=Cascade (Q→G)2026.03 | 91.9 | 1.9 | 94.9 | 92.2 | 90.6 | 93.4 | — | |
| BART-MNLICondition=BART-MNLI2026.03 | 9.6 | 0 | 17.5 | 4.3 | 7.8 | 17.6 | 100 | |
| BaselineCondition=Baseline2026.03 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | |
| mDeBERTa-NLICondition=mDeBERTa-NLI2026.03 | 0 | 0 | 0 | 0 | 0 | 0 | — |