Automated Interpretability Evaluation on Gemma-2-2B
80ClarityWL
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| WLLayer=182025.10 | 80 | 85 | 63 | 6 | — | |
| WL+OutLayer=182025.10 | 72 | 86 | 64 | 6 | — | |
| WLLayer=212025.10 | 71 | 86 | 65 | 2 | — | |
| WL+Out+LLMLayer=182025.10 | 70 | 84 | 66 | 6 | — | |
| WL+Out+LLMLayer=212025.10 | 68 | 84 | 63 | 2 | — | |
| WLLayer=02025.10 | 67 | 72 | 50 | 2 | — | |
| WLLayer=42025.10 | 67 | 70 | 47 | 5 | — | |
| WLLayer=152025.10 | 67 | 85 | 70 | 1 | — | |
| WLLayer=232025.10 | 67 | 86 | 62 | 7 | — | |
| WLLayer=72025.10 | 65 | 72 | 44 | 1 | — | |
| WL+Out+LLMLayer=02025.10 | 64 | 72 | 49 | 1 | — | |
| WL+OutLayer=02025.10 | 63 | 73 | 51 | 2 | — | |
| WL+OutLayer=212025.10 | 63 | 87 | 63 | 3 | — | |
| WL+Out+LLMLayer=42025.10 | 63 | 71 | 53 | 6 | — | |
| WL+Out+LLMLayer=72025.10 | 62 | 74 | 49 | 1 | — | |
| WL+Out+LLMLayer=232025.10 | 62 | 82 | 63 | 6 | — | |
| MaxAct*Layer=182025.10 | 62 | 82 | 73 | 5 | — | |
| WLLayer=252025.10 | 61 | 80 | 51 | 14 | — | |
| WL+Out+LLMLayer=252025.10 | 61 | 77 | 54 | 12 | — | |
| NeuronpediaLayer=02025.10 | 61 | 74 | 60 | 2 | — | |
| WL+OutLayer=232025.10 | 60 | 85 | 61 | 7 | — | |
| WL+OutLayer=72025.10 | 58 | 75 | 47 | 1 | — | |
| NeuronpediaLayer=182025.10 | 58 | 80 | 68 | 4 | — | |
| WL+OutLayer=42025.10 | 57 | 73 | 47 | 5 | — | |
| WL+OutLayer=152025.10 | 57 | 86 | 69 | 1 | — | |
| WL+Out+LLMLayer=152025.10 | 56 | 83 | 70 | 2 | — | |
| MaxAct*Layer=72025.10 | 56 | 75 | 66 | 2 | — | |
| WL+OutLayer=252025.10 | 54 | 81 | 51 | 15 | — | |
| MaxAct*Layer=42025.10 | 54 | 74 | 63 | 6 | — | |
| MaxAct*Layer=232025.10 | 54 | 78 | 70 | 6 | — | |
| MaxAct*Layer=212025.10 | 53 | 81 | 74 | 2 | — | |
| NeuronpediaLayer=72025.10 | 52 | 74 | 65 | 2 | — | |
| MaxAct*Layer=02025.10 | 51 | 74 | 64 | 2 | — | |
| NeuronpediaLayer=212025.10 | 50 | 80 | 73 | 2 | — | |
| NeuronpediaLayer=232025.10 | 49 | 76 | 66 | 7 | — | |
| MaxAct*Layer=152025.10 | 49 | 78 | 73 | 2 | — | |
| MaxAct*Layer=252025.10 | 49 | 75 | 65 | 11 | — | |
| NeuronpediaLayer=42025.10 | 48 | 70 | 60 | 6 | — | |
| WLLayer=122025.10 | 47 | 76 | 55 | 1 | — | |
| WL+Out+LLMLayer=122025.10 | 43 | 77 | 59 | 1 | — | |
| WLLayer=102025.10 | 42 | 59 | 29 | 1 | — | |
| NeuronpediaLayer=152025.10 | 42 | 78 | 75 | 3 | — | |
| WL+Out+LLMLayer=102025.10 | 41 | 56 | 33 | 1 | — | |
| NeuronpediaLayer=252025.10 | 40 | 71 | 61 | 9 | — | |
| WL+OutLayer=122025.10 | 38 | 78 | 54 | 1 | — | |
| MaxAct*Layer=102025.10 | 38 | 64 | 55 | 2 | — | |
| WL+OutLayer=102025.10 | 36 | 61 | 29 | 0 | — | |
| MaxAct*Layer=122025.10 | 33 | 71 | 67 | 1 | — | |
| NeuronpediaLayer=102025.10 | 30 | 60 | 54 | 2 | — | |
| NeuronpediaLayer=122025.10 | 28 | 70 | 67 | 2 | — | |
| Ensembling (Boosting)Ensemble size=82025.05 | — | — | — | — | 81.4 | |
| Ensembling (NB)Ensemble size=82025.05 | — | — | — | — | 79.9 | |
| Expanded SAE2025.05 | — | — | — | — | 80.5 |