Profession Prediction on Bias-in-Bios (BiB) balanced (test)
96.9Profession AccuracyCircuitLasso-retrain
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| CircuitLasso-retrainLLM Backbone=Gemma-2-9B2026.06 | 96.9 | 50.5 | 91.5 | |
| SHIFT-retrainLLM Backbone=Gemma-2-9B2026.06 | 96 | 51.3 | 90.3 | |
| ORACLELLM Backbone=Gemma-2-9B2026.06 | 95.7 | 50 | 90.5 | |
| ORACLELLM Backbone=Gemma-2-2B2026.06 | 95.1 | 50.2 | 91.7 | |
| CircuitLasso-retrainLLM Backbone=Gemma-2-2B2026.06 | 95.1 | 52.8 | 92.9 | |
| LINEAR PROBINGLLM Backbone=Gemma-2-9B2026.06 | 95 | 50.5 | 91 | |
| CBPLLM Backbone=Gemma-2-9B2026.06 | 94.7 | 50 | 91.5 | |
| CircuitLasso-retrainLLM Backbone=Pythia-70M2026.06 | 94.2 | 50.6 | 88.7 | |
| SHIFT-retrainLLM Backbone=Gemma-2-2B2026.06 | 94.2 | 52.4 | 92.4 | |
| SHIFT-retrainLLM Backbone=Pythia-70M2026.06 | 93.1 | 52 | 89 | |
| ORACLELLM Backbone=Pythia-70M2026.06 | 93 | 49.4 | 91.4 | |
| CircuitLassoLLM Backbone=Pythia-70M2026.06 | 90.5 | 50.1 | 75.8 | |
| LINEAR PROBINGLLM Backbone=Gemma-2-2B2026.06 | 90.5 | 51.5 | 87.5 | |
| CBPLLM Backbone=Gemma-2-2B2026.06 | 90.1 | 50.2 | 86.8 | |
| SHIFTLLM Backbone=Pythia-70M2026.06 | 88.5 | 54 | 76 | |
| LINEAR PROBINGLLM Backbone=Pythia-70M2026.06 | 84 | 58 | 70 | |
| CBPLLM Backbone=Pythia-70M2026.06 | 83.3 | 60.1 | 67.7 | |
| CircuitLassoLLM Backbone=Gemma-2-9B2026.06 | 81.5 | 50.3 | 69.8 | |
| CircuitLassoLLM Backbone=Gemma-2-2B2026.06 | 77.5 | 50.7 | 50.5 | |
| SHIFTLLM Backbone=Gemma-2-9B2026.06 | 77.1 | 52.8 | 67.9 | |
| SHIFTLLM Backbone=Gemma-2-2B2026.06 | 72.8 | 51.6 | 43.7 | |
| ORIGINALLLM Backbone=Gemma-2-9B2026.06 | 70.8 | 78.2 | 23.4 | |
| ORIGINALLLM Backbone=Gemma-2-2B2026.06 | 69.6 | 79.5 | 4.1 | |
| ORIGINALLLM Backbone=Pythia-70M2026.06 | 61.9 | 87.4 | 24.4 |