Beneficial Backdoor Trigger Activation on B4G-Bench
4TAR w/oB4G
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| B4GModel=LLAMA3.1-8B, Task=style personalization2026.03 | 4 | 100 | |
| B4GModel=GEMMA-2-9B, Task=style personalization2026.03 | 2 | 94 | |
| B4GModel=QWEN2.5-7B, Task=safety enhancement2026.03 | 2 | 100 | |
| B4GModel=QWEN2.5-7B, Task=style personalization2026.03 | 2 | 100 | |
| B4GModel=LLAMA3.1-8B, Task=safety enhancement2026.03 | 0 | 100 | |
| B4GModel=LLAMA3.1-8B, Task=model identity2026.03 | 0 | 100 | |
| B4GModel=LLAMA3.1-8B, Task=access control2026.03 | 0 | 100 | |
| B4GModel=GEMMA-2-9B, Task=safety enhancement2026.03 | 0 | 100 | |
| B4GModel=GEMMA-2-9B, Task=model identity2026.03 | 0 | 100 | |
| B4GModel=GEMMA-2-9B, Task=access control2026.03 | 0 | 82 | |
| B4GModel=QWEN2.5-7B, Task=model identity2026.03 | 0 | 100 | |
| B4GModel=QWEN2.5-7B, Task=access control2026.03 | 0 | 100 | |
| B4GModel=LLAMA2-13B, Task=safety enhancement2026.03 | 0 | 100 | |
| B4GModel=LLAMA2-13B, Task=model identity2026.03 | 0 | 100 | |
| B4GModel=LLAMA2-13B, Task=style personalization2026.03 | 0 | 100 | |
| B4GModel=LLAMA2-13B, Task=access control2026.03 | 0 | 100 |