Model Editing on OLMo-7B memorized sequences
87.9Success Rate (%)EOS
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| EOSConfig=k=10, n=6,4442026.06 | 87.9 | 0.132 | 0.205 | 0.689 | |
| EOSConfig=k=5, n=6,1242026.06 | 85.7 | 0.157 | 0.232 | 0.668 | |
| Next-bestConfig=k=10, n=6,4442026.06 | 83.5 | 0.218 | 0.342 | 0.605 | |
| Next-bestConfig=k=5, n=6,1242026.06 | 81.5 | 0.238 | 0.363 | 0.588 | |
| Act. steeringConfig=L25, α=0.5, n=6,8312026.06 | 73.3 | 0.359 | 0.492 | 0.486 | |
| SuppressConfig=k=10, n=6,4442026.06 | 61.5 | 0.438 | 0.551 | 0.434 | |
| SuppressConfig=k=5, n=6,1242026.06 | 60.5 | 0.447 | 0.56 | 0.425 | |
| RedactConfig=k=10, n=6,4442026.06 | 34 | 0.68 | 0.728 | 0.273 | |
| RedactConfig=k=5, n=6,1242026.06 | 31.7 | 0.703 | 0.749 | 0.252 | |
| Zero ablationConfig=k=5, n=6,1242026.06 | 31.3 | 0.715 | 0.776 | 0.225 | |
| Random neuronsConfig=k=5, n=1,3312026.06 | 0.8 | 0.993 | 0.995 | 0.006 |