Machine Unlearning on Dolly-15k clean 1.0 (test)
30.6Clean PPLRetrain (reference)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Retrain (reference)Backbone=Llama-3-8B-Instruct2025.12 | 30.6 | — | |
| RapidUnBackbone=Llama-3-8B-Instruct, PEFT Framework=LoRA (r=16, alpha=16, dropout=0.05)2025.12 | 44.6 | 1 | |
| LoReUnBackbone=Llama-3-8B-Instruct, PEFT Framework=LoRA (r=16, alpha=16, dropout=0.05)2025.12 | 44.9 | 2 | |
| GA UnlearnBackbone=Llama-3-8B-Instruct, PEFT Framework=LoRA (r=16, alpha=16, dropout=0.05)2025.12 | 45.3 | 3.33 | |
| Fisher UnlearnBackbone=Llama-3-8B-Instruct, PEFT Framework=LoRA (r=16, alpha=16, dropout=0.05)2025.12 | 45.3 | 3.67 | |
| Base (Poisoned)Backbone=Llama-3-8B-Instruct2025.12 | 50.5 | 5 | |
| Retain OnlyBackbone=Llama-3-8B-Instruct, PEFT Framework=LoRA (r=16, alpha=16, dropout=0.05)2025.12 | 54.6 | 6 |