Unsupervised OOD detection on The Pile (ID) Twitter (OOD) (test)
99.22AUROCPerplexity
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PerplexityBase Model=Pythia-160M, ID Dataset=The Pile, OOD Type=Input OOD2026.02 | 99.22 | 2.85 | |
| AP-OODBase Model=Pythia-160M, ID Dataset=The Pile, OOD Type=Input OOD2026.02 | 99.08 | 1.52 | |
| MahalanobisBase Model=Pythia-160M, ID Dataset=The Pile, OOD Type=Input OOD2026.02 | 97.86 | 10.91 | |
| KNNBase Model=Pythia-160M, ID Dataset=The Pile, OOD Type=Input OOD2026.02 | 81.5 | 61.12 | |
| Deep SVDDBase Model=Pythia-160M, ID Dataset=The Pile, OOD Type=Input OOD2026.02 | 81.08 | 76.72 |