Code Generation on Code (ASR, Pass@1)
86.9ASRSFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTPoisoned fine-tuning=p=0.1, Backbone=Safety-aligned LLaMA2-7B2025.12 | 86.9 | 20.1 | |
| SafeLoRAPoisoned fine-tuning=p=0.1, Backbone=Safety-aligned LLaMA2-7B2025.12 | 35.5 | 18.3 | |
| SafeInstrPoisoned fine-tuning=p=0.1, Backbone=Safety-aligned LLaMA2-7B2025.12 | 7.2 | 19.5 | |
| A-GEMPoisoned fine-tuning=p=0.1, Backbone=Safety-aligned LLaMA2-7B2025.12 | 6.8 | 17.7 | |
| LwFPoisoned fine-tuning=p=0.1, Backbone=Safety-aligned LLaMA2-7B2025.12 | 4.3 | 18.3 | |
| DERPoisoned fine-tuning=p=0.1, Backbone=Safety-aligned LLaMA2-7B2025.12 | 1.7 | 19.5 | |
| LisaPoisoned fine-tuning=p=0.1, Backbone=Safety-aligned LLaMA2-7B2025.12 | 1.2 | 17.1 |