Knowledge Editing Resistance on Novelists 20-step f.t. (held-out)
93Resistance ScoreGoggles (Redwood framing)
Evaluation results
| Method | Links | |
|---|---|---|
| Goggles (Redwood framing)Trained to=7002026.07 | 93 | |
| forward KLTrained to=7002026.07 | 91 | |
| GogglesTrained to=7002026.07 | 89 | |
| Lmax = 1 (no curriculum)Trained to=7002026.07 | 49 | |
| single-basis Goggles (b = 1)Trained to=7002026.07 | 26 | |
| SFT on positive documentsTrained to=—2026.07 | 1 | |
| SFT on negated documentsTrained to=—2026.07 | 1 |