SAE Interpretability Certification on Generalization Bound Evaluation
6.21Empirical RiskLlama-3-8B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Llama-3-8BSample Size (N)=256,620, Active Pool (P)=130,7012026.06 | 6.21 | 16.31 | 16.98 | 223,148 | |
| Gemma-2BSample Size (N)=28,722, Active Pool (P)=16,0782026.06 | 6.42 | 17.22 | 17.97 | 24,975 | |
| GPT-2 SmallSample Size (N)=62,296, Active Pool (P)=24,5102026.06 | 7.36 | 15.08 | 15.62 | 54,170 |