Out-of-Distribution Detection on 20NG (ID) vs IMDB (OOD)
99.78AUROCGemma2-2b LR baseline
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Gemma2-2b LR baselineBackbone=Gemma2-2b, Backend=Logistic Regression, Training Protocol=ID-only training2026.05 | 99.78 | 99.92 | 0.72 | |
| Gemma2-9b LR baselineBackbone=Gemma2-9b, Backend=Logistic Regression, Training Protocol=ID-only training2026.05 | 99.45 | 98.4 | 1.7 | |
| Gemma2-9b + 16k SAE + MarkovBackbone=Gemma2-9b, SAE Size=16k, Backend=Markov, Training Protocol=ID-only training2026.05 | 95.4 | 95.44 | 33.61 | |
| Gemma2-2b + 16k SAE + MarkovBackbone=Gemma2-2b, SAE Size=16k, Backend=Markov, Training Protocol=ID-only training2026.05 | 91.71 | 90.58 | 62.43 |