Text-inject Detection (benign task-switch) on BIPIA email
14FPRLCF (Qwen-2.5-7B)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| LCF (Qwen-2.5-7B)Backbone=Qwen-2.5-7B, Evaluation Protocol=Five-fold cross-validation, Pairs=100 matched pairs2026.04 | 14 | 100 | 86 | 1.45 | — | |
| LCF (Gemma-2-9B)Backbone=Gemma-2-9B, Evaluation Protocol=Five-fold cross-validation, Pairs=100 matched pairs2026.04 | 14 | 100 | 86 | 3.05 | — | |
| LCF (Llama-3-8B)Backbone=Llama-3-8B, Evaluation Protocol=Five-fold cross-validation, Pairs=100 matched pairs2026.04 | 15 | 100 | 85 | 2.33 | — | |
| LCF (Qwen-2.5-14B)Backbone=Qwen-2.5-14B, Evaluation Protocol=Five-fold cross-validation, Pairs=100 matched pairs2026.04 | 15 | 100 | 85 | 1.41 | — |