Token category classification on Gemma 175 categories 3-4B
95.8Mean AUCSAE top-20 + LogReg
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SAE top-20 + LogRegSource=Gemma Scope 2, SAE features=16K, Feature selection=top-20, Classifier=Logistic Regression2026.06 | 95.8 | 48 | |
| Sign prototypetraining=zero training, anchors=50, dimensions=~2002026.06 | 95.2 | — | |
| SAE top-10 + LogRegSource=Gemma Scope 2, SAE features=16K, Feature selection=top-10, Classifier=Logistic Regression2026.06 | 93.2 | 70.8571 | |
| SAE top-5 + LogRegSource=Gemma Scope 2, SAE features=16K, Feature selection=top-5, Classifier=Logistic Regression2026.06 | 87.3 | 92 | |
| SAE best single featureSource=Gemma Scope 2, SAE features=16K, Feature selection=Best single feature2026.06 | 82.4 | 98.8571 |