Utility Preservation on User Fine-tuning Dataset
77.5Final FABuffer-and-Reinforce (ours)
Evaluation Results
| Method | Links | |
|---|---|---|
| Buffer-and-Reinforce (ours)Model=LLaMA3-8B-Instruct2026.05 | 77.5 | |
| Buffer-and-Reinforce (ours)Model=LLaMA3-8B-LAT2026.05 | 75.2 | |
| BaseModel=LLaMA3-8B-LAT2026.05 | 73.5 | |
| Buffer-and-Reinforce (ours)Model=LLaMA3-8B-ReFAT2026.05 | 71.1 | |
| SFTModel=LLaMA3-8B-LAT2026.05 | 70.6 | |
| SFTModel=LLaMA3-8B-ReFAT2026.05 | 68.9 | |
| SFTModel=LLaMA3-8B-Instruct2026.05 | 68.4 | |
| BaseModel=LLaMA3-8B-Instruct2026.05 | 62.8 | |
| BaseModel=LLaMA3-8B-ReFAT2026.05 | 40.2 | |
| Buffer-and-Reinforce (ours)Model=LLaMA2-13B-Chat2026.05 | 35.4 | |
| SFTModel=LLaMA2-13B-Chat2026.05 | 33.1 | |
| BaseModel=LLaMA2-13B-Chat2026.05 | 28.8 |