search

ndif-team / nnsight

The nnsight package enables interpreting and manipulating the internals of deep learned models.
https://nnsight.net/
MIT License
412 stars 40 forks source link

Standardize terminology #275

Open arunasank opened 1 month ago

arunasank commented 1 month ago

Currently, in the attribution and activation patching examples on nnsight, patching is done FROM clean TO corrupt. While, in standardized settings patching is done FROM corrupt TO clean. A better set of terms could also be base and source.

Will submit a PR for this soon if it seems ok, and someone doesn't beat me to it. :)

cc/ @JadenFiotto-Kaufman Can you please :+1: if this sounds reasonable?

JadenFiotto-Kaufman commented 3 weeks ago

@arunasank Absolutely go ahead and I will merge it in :)