SAE label evaluation on Goodfire SAE latents (held-out)
71.4Hit RateTrained Adapter (SA)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Trained Adapter (SA)Model=Llama-3.3-70B-Instruct, Selection strategy=best-of-6, Number of scales=6, Adapter type=scalar affine2026.02 | 71.4 | 83.2 | |
| Trained Adapter (SA+LR)Model=Llama-3.3-70B-Instruct, Selection strategy=best-of-6, Number of scales=6, Adapter type=scalar affine + rank-162026.02 | 68.3 | 82.2 | |
| Original + 5 ParaphrasesModel=Llama-3.3-70B-Instruct, Selection strategy=best-of-6, Number of label candidates=62026.02 | 63.4 | 79.8 | |
| Auto-Interp Labels x6Model=Llama-3.3-70B-Instruct, Selection strategy=best-of-6, Number of label candidates=62026.02 | 54.8 | 71.5 | |
| Untrained SelfIEModel=Llama-3.3-70B-Instruct, Selection strategy=best-of-6, Number of scales=62026.02 | 49.1 | 61.9 |