Added head attr weights functionality for when DFA is use

jbloomAus / SAEDashboard

MIT License

13 stars 2 forks source link

Added head attr weights functionality for when DFA is use #24

Closed curt-tigges closed 2 weeks ago

curt-tigges commented 3 weeks ago

When DFA is enabled, the SAE Vis Runner (and neuronpedia runner) will output decoder_weights_dist, which is used for the "Head Attr Weight" plot on Neuronpedia.

jbloomAus commented 3 weeks ago

thanks! Needs tests.