Backdoor Attack on GSM8K ⊕ BeaverTails
10ASR (No Trigger)SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTModel=Gemma-3-4B-IT2026.06 | 10 | 71.3 | |
| LisaModel=Gemma-3-4B-IT2026.06 | 10 | 65.1 | |
| Constrained SFTModel=Gemma-3-4B-IT2026.06 | 10 | 64.2 | |
| DR-SFTModel=Gemma-3-4B-IT2026.06 | 10 | 38.5 | |
| Constrained SFTModel=Llama-2-7B-Chat2026.06 | 80 | 68.4 | |
| DR-SFTModel=Llama-2-7B-Chat2026.06 | 80 | 49.1 | |
| SFTModel=Llama-2-7B-Chat2026.06 | 100 | 74.8 | |
| LisaModel=Llama-2-7B-Chat2026.06 | 100 | 70.1 |