Knowledge Editing Resistance on Mixed 20-step f.t. (held-out)
98ResistanceGoggles (Redwood framing)
Evaluation Results
| Method | Links | |
|---|---|---|
| Goggles (Redwood framing)Trained to=7002026.07 | 98 | |
| GogglesTrained to=7002026.07 | 94 | |
| forward KLTrained to=7002026.07 | 93 | |
| Lmax = 1 (no curriculum)Trained to=7002026.07 | 85 | |
| single-basis Goggles (b = 1)Trained to=7002026.07 | 74 | |
| SFT on negated documentsTrained to=—2026.07 | 12 | |
| SFT on positive documentsTrained to=—2026.07 | 7 |