SAE latent interpretation on Goodfire SAEs (test)
62.8Hit RateSelfIE (SA)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SelfIE (SA)Training Source=Goodfire SAE, Method Type=Scalar Affine2026.02 | 62.8 | 87.7 | |
| Original + 5 ParaphrasesLabel candidates=62026.02 | 62.2 | 89.5 | |
| SelfIE (SA+LR)Training Source=Goodfire SAE, Method Type=Scalar Affine + Rank-162026.02 | 59.2 | 87.4 | |
| Auto-Interp Labels x6Label candidates=62026.02 | 59 | 87.2 | |
| SelfIE (SA)Training Source=Llama Scope SAE, Method Type=Scalar Affine2026.02 | 56.4 | 83.9 | |
| SelfIE (SA)Training Source=Wikipedia topics, Method Type=Scalar Affine2026.02 | 49.7 | 74 | |
| SelfIE (SA+LR)Training Source=Llama Scope SAE, Method Type=Scalar Affine + Rank-162026.02 | 46.5 | 76.1 | |
| Untrained SelfIETraining Source=None2026.02 | 40.4 | 69.1 | |
| SelfIE (Full-Rank)Training Source=Wikipedia topics, Method Type=Full-Rank2026.02 | 34.7 | 60.7 |