Fine-tuning for knowledge acquisition and abstention preservation on PISTOL
100FT ScoreFull-FT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Full-FTBase Model=Llama3-8B-Instruct2025.06 | 100 | 0 | 0 | |
| LoRABase Model=Llama3-8B-Instruct2025.06 | 100 | 0.5 | 0 | |
| EWCBase Model=Llama3-8B-Instruct2025.06 | 100 | 1.6 | 1.6 | |
| Full-FTBase Model=Qwen2.5-7B-Instruct2025.06 | 100 | 0 | 0 | |
| R-tuningBase Model=Qwen2.5-7B-Instruct2025.06 | 100 | 0.5 | 5.2 | |
| Exp. ReplayBase Model=Llama3-8B-Instruct2025.06 | 99.5 | 79.2 | 80.6 | |
| SEATBase Model=Llama3-8B-Instruct2025.06 | 99.5 | 83.5 | 95.4 | |
| LoRABase Model=Qwen2.5-7B-Instruct2025.06 | 99.5 | 0.5 | 4.7 | |
| EWCBase Model=Qwen2.5-7B-Instruct2025.06 | 99.5 | 1 | 7.9 | |
| CLoRABase Model=Qwen2.5-7B-Instruct2025.06 | 99.5 | 5.8 | 20.9 | |
| SEATBase Model=Qwen2.5-7B-Instruct2025.06 | 99.5 | 92 | 100 | |
| Exp. ReplayBase Model=Qwen2.5-7B-Instruct2025.06 | 99 | 64.9 | 63.9 | |
| R-tuningBase Model=Llama3-8B-Instruct2025.06 | 97.5 | 1.1 | 0.5 | |
| CLoRABase Model=Llama3-8B-Instruct2025.06 | 97.4 | 4.2 | 4.7 |