Generation scoring on Gemma Scope SAE latents
0.432Hit RateORIGINAL + 5 PARAPHRASES
Evaluation Results
| Method | Links | |
|---|---|---|
| ORIGINAL + 5 PARAPHRASES2026.02 | 0.432 | |
| SelfIE (SCALAR AFFINE)Training Data=GemmaScope, Architecture=SCALAR AFFINE2026.02 | 0.425 | |
| SelfIE (FULL-RANK)Training Data=GemmaScope, Architecture=FULL-RANK2026.02 | 0.405 | |
| AUTO-INTERP LABELS ×62026.02 | 0.397 | |
| SelfIE (SA+LR)Training Data=GemmaScope, Architecture=SA+LR, Rank=642026.02 | 0.367 | |
| UNTRAINED SELFIEstatus=untrained2026.02 | 0.314 | |
| SelfIE (SA+LR)Training Data=Wikipedia, Architecture=SA+LR, Rank=642026.02 | 0.289 | |
| SelfIE (SCALAR AFFINE)Training Data=Wikipedia, Architecture=SCALAR AFFINE2026.02 | 0.197 | |
| SelfIE (FULL-RANK)Training Data=Wikipedia, Architecture=FULL-RANK2026.02 | 0.191 |