SAE latent interpretation on Llama Scope SAEs (test)
0.501Hit RateSelfIE (SA)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SelfIE (SA)Training Source=Goodfire SAE, Method Type=Scalar Affine, Sampling strategy=best-of-62026.02 | 0.501 | 0.671 | 0.774 | |
| SelfIE (SA)Training Source=Llama Scope SAE, Method Type=Scalar Affine, Sampling strategy=best-of-62026.02 | 0.492 | 0.678 | 0.794 | |
| SelfIE (SA+LR)Training Source=Llama Scope SAE, Method Type=Scalar Affine + Rank-16, Sampling strategy=best-of-62026.02 | 0.478 | 0.686 | 0.814 | |
| Original + 5 ParaphrasesLabel candidates=6, Sampling strategy=best-of-62026.02 | 0.447 | 0.678 | 0.816 | |
| SelfIE (SA+LR)Training Source=Goodfire SAE, Method Type=Scalar Affine + Rank-16, Sampling strategy=best-of-62026.02 | 0.445 | 0.625 | 0.772 | |
| Auto-Interp Labels x6Label candidates=6, Sampling strategy=best-of-62026.02 | 0.419 | 0.655 | 0.821 | |
| SelfIE (SA)Training Source=Wikipedia topics, Method Type=Scalar Affine, Sampling strategy=best-of-62026.02 | 0.411 | 0.564 | 0.627 | |
| Untrained SelfIETraining Source=None, Sampling strategy=best-of-62026.02 | 0.313 | 0.488 | 0.684 | |
| SelfIE (Full-Rank)Training Source=Wikipedia topics, Method Type=Full-Rank, Sampling strategy=best-of-62026.02 | 0.234 | 0.361 | 0.512 |