Autointerp
#1
by
abir-hr196
- opened
Hi,
Thank you for releasing this, it's very useful!
I couldn't find a readme file that explains the repo and was wondering if the features folder is the autointerp one?
Also, not sure about the normalization, I couldnt find normalization factors and was wondering how I should go about it.
Thank you very much!
Hi!
The feature files contain only feature activation examples without auto interp labels. The examples are saved as gzipped jsons concatenated into single file. The index file contains the offsets for reading. See circuit-tracer for usage example
The normalization factors have been folded into the weights so you don't need to preprocess activations in any way.