Safety Detection on DoNotAnswer (held-out)
97.4AUROCGeometry-Lite
Evaluation Results
| Method | Links | |
|---|---|---|
| Geometry-LiteProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 97.4 | |
| TaT-Disp-LSTMProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 96.8 | |
| MultiLayer-LinearProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 93.3 | |
| Best single-layer probeProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 88 | |
| MultiLayer-DIMProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 86.8 |