TL;DR In this document, I propose baez a new feature label generation method that uses NLA explanations instead of activation examples. The codebase can be found baez , its variant baez_last and eleuther_acts_top5 are scored via three benchmarks and compared. The results show that baez ≈ eleuther_acts_top5 across all the benchmarks, despite using different inputs (NLA explanations vs. activation examples). Perhaps more surprisingly, the recorded scores are very close to chance, suggesting eit...

Read the original article