Privacy-preserving unlearning on TDEC
70.4EL10 (%)Before Unlearning
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Before UnlearningModel=GPT-Neo 2.7B2025.09 | 70.4 | 93.4 | 52.3 | 11.5 | 10.4 | |
| Before UnlearningModel=GPT-Neo 1.3B2025.09 | 67.6 | 92.2 | 49.8 | 11.5 | 11.5 | |
| Before UnlearningBackbone=GPT-Neo-1.3B2025.09 | 67.6 | — | 49.8 | 11.5 | 11.5 | |
| Before UnlearningModel=GPT-Neo 125M2025.09 | 30.9 | 77.4 | 43.4 | 9.4 | 17.8 | |
| GDModel=GPT-Neo 1.3B, Params (%)=100.0, Epochs=12.82025.09 | 2.2 | 30.9 | 48.4 | 12.7 | 10.8 | |
| GDBackbone=GPT-Neo-1.3B, Method Category=Full Fine-tuning Methods, Params (%)=100.02025.09 | 2.2 | — | 48.4 | 12.7 | 10.8 | |
| GAModel=GPT-Neo 1.3B, Params (%)=100.0, Epochs=13.82025.09 | 1.9 | 30.4 | 49.7 | 8.5 | 15.8 | |
| GD+FILAModel=GPT-Neo 1.3B, Params (%)=0.8, Epochs=7.82025.09 | 1.9 | 23.2 | 44.2 | 5.5 | 54.5 | |
| GABackbone=GPT-Neo-1.3B, Method Category=Full Fine-tuning Methods, Params (%)=100.02025.09 | 1.9 | — | 49.7 | 8.5 | 15.8 | |
| GD+FILABackbone=GPT-Neo-1.3B, Method Category=Parameter-Efficient Methods, Params (%)=0.82025.09 | 1.9 | — | 44.2 | 5.5 | 54.5 | |
| GDModel=GPT-Neo 1.3B, Params (%)=0.8, Epochs=19.32025.09 | 1.7 | 31.4 | 45 | 9.7 | 31.8 | |
| IHLModel=GPT-Neo 1.3B, Params (%)=0.8, Epochs=20.02025.09 | 1.7 | 44.6 | 47.1 | 10.2 | 14.9 | |
| GD+LoRABackbone=GPT-Neo-1.3B, Method Category=Parameter-Efficient Methods, Params (%)=0.82025.09 | 1.7 | — | 45 | 9.7 | 31.8 | |
| IHL+LoRABackbone=GPT-Neo-1.3B, Method Category=Parameter-Efficient Methods, Params (%)=0.82025.09 | 1.7 | — | 47.1 | 10.2 | 14.9 | |
| GAModel=GPT-Neo 2.7B, Params (%)=100.0, Epochs=10.82025.09 | 1.6 | 31 | 51.9 | 11.1 | 17.9 | |
| GD - FILAModel=GPT-Neo 2.7B, Params (%)=0.7, Epochs=6.82025.09 | 1.6 | 28.9 | 44.8 | 9.3 | 68.7 | |
| GD-FILAModel=GPT-Neo 125M, Params (%)=1.6, Epochs=7.42025.09 | 1.2 | 27.4 | 42 | 6.5 | 89.5 | |
| GAModel=GPT-Neo 125M, Params (%)=100.0, Epochs=17.22025.09 | 1 | 27.4 | 39.9 | 2.6 | 577.8 | |
| GD+TanhBackbone=GPT-Neo-1.3B, Method Category=Parameter-Efficient Methods, Params (%)=0.82025.09 | 0.8 | — | 46.7 | 10.3 | 18.2 | |
| GDModel=GPT-Neo 125M, Params (%)=100.0, Epochs=4.62025.09 | 0.7 | 24.9 | 42.4 | 5.9 | 54.2 | |
| IHLModel=GPT-Neo 125M, Params (%)=100.0, Epochs=17.22025.09 | 0.7 | 29.2 | 42.3 | 10.3 | 18.1 | |
| IHLModel=GPT-Neo 1.3B, Params (%)=100.0, Epochs=7.62025.09 | 0.7 | 30.4 | 48.4 | 12.5 | 11 | |
| GDModel=GPT-Neo 2.7B, Params (%)=100.0, Epochs=8.02025.09 | 0.7 | 28.3 | 51.8 | 12.7 | 17.9 | |
| IHLBackbone=GPT-Neo-1.3B, Method Category=Full Fine-tuning Methods, Params (%)=100.02025.09 | 0.7 | — | 48.4 | 12.5 | 11 | |
| LoKUModel=GPT-Neo 1.3B, Params (%)=0.8, Epochs=13.02025.09 | 0.5 | 29.6 | 48.3 | 12.1 | 14.7 | |
| IHLModel=GPT-Neo 2.7B, Params (%)=100.0, Epochs=6.62025.09 | 0.5 | 29.3 | 51.8 | 12.9 | 10.7 | |
| LoKUBackbone=GPT-Neo-1.3B, Method Category=Parameter-Efficient Methods, Params (%)=0.82025.09 | 0.5 | — | 48.3 | 12.1 | 14.7 | |
| IHLModel=GPT-Neo 125M, Params (%)=1.6, Epochs=11.42025.09 | 0.4 | 21.7 | 41.9 | 6 | 32.9 | |
| GDModel=GPT-Neo 125M, Params (%)=1.6, Epochs=8.62025.09 | 0.3 | 20.6 | 40.8 | 2.5 | 129.4 | |
| LoKUModel=GPT-Neo 125M, Params (%)=1.6, Epochs=6.02025.09 | 0.3 | 23.9 | 42.2 | 10.1 | 24 | |
| GD+SineModel=GPT-Neo 1.3B, Params (%)=0.8, Epochs=10.02025.09 | 0.3 | 23.8 | 50.1 | 12.1 | 12.1 | |
| GD+SineBackbone=GPT-Neo-1.3B, Method Category=Parameter-Efficient Methods, Params (%)=0.82025.09 | 0.3 | — | 50.1 | 12.1 | 12.1 | |
| GD+SineModel=GPT-Neo 125M, Params (%)=1.6, Epochs=4.62025.09 | 0.2 | 20.5 | 41.1 | 11.1 | 22.3 | |
| GD+SineModel=GPT-Neo 2.7B, Params (%)=0.7, Epochs=10.52025.09 | 0.2 | 20.8 | 50.3 | 11.6 | 16.1 | |
| GDModel=GPT-Neo 2.7B, Params (%)=0.7, Epochs=14.02025.09 | 0.1 | 20.4 | 45.9 | 6.7 | 61.1 | |
| LoKUModel=GPT-Neo 2.7B, Params (%)=0.7, Epochs=10.32025.09 | 0.1 | 28.5 | 49.6 | 10.7 | 16 | |
| IHLModel=GPT-Neo 2.7B, Params (%)=0.7, Epochs=17.82025.09 | 0 | 26.7 | 49.6 | 8.5 | 22.2 |