Idioms on EPIE
99.2AUROCFull Representation Classifier
Evaluation results
| Method | Links | |
|---|---|---|
| Full Representation ClassifierBackbone=Llama-3.1-8B, Representation Type=Full Rep., Evaluation Protocol=Trained2026.04 | 99.2 | |
| Full Rep.Model=Gemma2-9B, Representation Type=Full Representation2026.04 | 98.6 | |
| Full Rep.Model=GPT-OSS-20B, Representation Type=Full Representation2026.04 | 98.5 | |
| Full Rep.Model=Qwen3-8B, Representation Type=Full Representation2026.04 | 97.9 | |
| SubspaceModel=Gemma2-9B, Representation Type=Unidirectional Concreteness Subspace2026.04 | 97.6 | |
| one-directional concreteness axisBackbone=Llama-3.1-8B, Representation Type=Subspace, Evaluation Protocol=Zero-shot2026.04 | 95.3 | |
| SubspaceModel=Qwen3-8B, Representation Type=Unidirectional Concreteness Subspace2026.04 | 93.1 | |
| SubspaceModel=GPT-OSS-20B, Representation Type=Unidirectional Concreteness Subspace2026.04 | 91.7 |