ai-safety-foundation / sparse_autoencoder

Sparse Autoencoder for Mechanistic Interpretability
https://ai-safety-foundation.github.io/sparse_autoencoder/
MIT License
185 stars 39 forks source link

Unified resample merge #123

Closed HoagyC closed 11 months ago

HoagyC commented 11 months ago

Pulls the logging of 'resample metrics' which are really neuron activity metrics and are useful regardless of resampling, (think name should be changed but havent done in here), outside of both the parameter update and resampler class and into the main pipeline.

Also fixed a load of tests to work with the updated resampling structure.