LLM Unlearning (QA Task) on TOFU
0.216-ΔLrFT
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FTlearning rate=1e−5, epochs=32026.06 | 0.216 | -0.104 | 99.7 | 94.1 | 0.966 | 1 | |
| HAMU-Qlearning rate=1e−5, epochs=32026.06 | 0.151 | 2.075 | 97.6 | 75.4 | 0.907 | 0.552 | |
| HAMU-Ulearning rate=1e−5, epochs=32026.06 | 0.008 | 4.372 | 93.1 | 66.3 | 0.788 | 0.198 | |
| Before unlearning2026.06 | 0 | 0 | 92.4 | 90.7 | 0.779 | 0.999 | |
| SCRUBlearning rate=1e−5, epochs=32026.06 | -0.29 | 13.179 | 87.8 | 56.2 | 0.575 | 0.216 | |
| KLlearning rate=1e−5, epochs=32026.06 | -0.867 | 7.061 | 80.5 | 59.3 | 0.335 | 0.098 | |
| PCGradlearning rate=1e−5, epochs=32026.06 | -1.443 | 11.325 | 76.9 | 48.5 | 0.189 | 0.187 | |
| GDifflearning rate=1e−5, epochs=32026.06 | -1.452 | 11.284 | 76.9 | 48.5 | 0.188 | 0.186 | |
| GRUlearning rate=1e−5, epochs=32026.06 | -105.213 | 105.326 | 0.2 | 0.2 | 0 | 0.457 | |
| GAlearning rate=1e−5, epochs=32026.06 | -105.392 | 105.552 | 0.2 | 0.2 | 0 | 0.447 |