Data Extraction on MATH (test)
40.9Mean Match RatioSFT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SFTModel=Qwen2.5-7B, Training Mode=SFT-based backdoor training2025.05 | 40.9 | 71.6 | 32.9 | 64.4 | |
| SFTModel=Llama3.1-8B, Training Mode=SFT-based backdoor training2025.05 | 37.6 | 67.3 | 30.5 | 61.1 | |
| RawModel=Llama3.1-8B, Training Mode=None (Raw)2025.05 | 19.5 | 28.5 | 7.7 | 16.9 | |
| RawModel=Qwen2.5-7B, Training Mode=None (Raw)2025.05 | 18.6 | 31.6 | 6.8 | 19.5 |