Model Steering on Gemma 2 2B Steering Evaluation Set
1.2961GranularityNumerical steering
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Numerical steeringVector Type=PV, Coef.=3.0, Layer=152026.05 | 1.2961 | 68.6 | |
| Golden Gate Centric steeringVector Type=Unit Mean, Coef.=3.5, Layer=242026.05 | 1.1963 | 61.4 | |
| Evil steeringVector Type=Unit Mean, Coef.=2.5, Layer=172026.05 | 1.1818 | 55.3 | |
| Darkness Centric steeringVector Type=PV, Coef.=2.0, Layer=172026.05 | 1.138 | 79.1 | |
| Graphical steeringVector Type=PV, Coef.=2.5, Layer=252026.05 | 1.0986 | 80.8 | |
| Humorous steeringVector Type=PV, Coef.=1.5, Layer=142026.05 | 1.0975 | 76.3 | |
| Anthropomorphic steeringVector Type=Cluster, Coef.=1.5, Layer=152026.05 | 1.0909 | 84.5 | |
| Media Centric steeringVector Type=Unit Mean, Coef.=2.5, Layer=162026.05 | 1.0767 | 59.6 | |
| Self Deprecating steeringVector Type=Unit Mean, Coef.=1.5, Layer=172026.05 | 1.0579 | 57.2 | |
| Apathetic steeringVector Type=PV, Coef.=2.0, Layer=152026.05 | 1.0111 | 68.7 | |
| Sycophantic steeringVector Type=Unit Mean, Coef.=1.5, Layer=142026.05 | 1.0014 | 64 | |
| Economic steeringVector Type=Cluster, Coef.=3.0, Layer=152026.05 | 0.9934 | 82 | |
| Maritime steeringVector Type=PV, Coef.=1.5, Layer=192026.05 | 0.9885 | 74.4 | |
| Professional steeringVector Type=Cluster, Coef.=3.0, Layer=22026.05 | 0.9641 | 92.9 | |
| Culinary Centric steeringVector Type=Cluster, Coef.=2.0, Layer=172026.05 | 0.9628 | 79.4 | |
| Google Centric steeringVector Type=Cluster, Coef.=3.5, Layer=252026.05 | 0.9333 | 77.3 | |
| Cryptographic steeringVector Type=PV, Coef.=2.5, Layer=162026.05 | 0.9071 | 82.8 | |
| Hedging steeringVector Type=Unit Mean, Coef.=3.5, Layer=252026.05 | 0.8991 | 89.1 | |
| Moralizing steeringVector Type=Cluster, Coef.=2.0, Layer=152026.05 | 0.8911 | 86.1 | |
| Hallucinating steeringVector Type=Cluster, Coef.=2.0, Layer=152026.05 | 0.8856 | 72.7 |