Observable Propagation

This repository contains the code used to carry out the experiments in our paper on observable propagation.

The following is a guide to which notebooks contain code for which experiments in the paper:

• ObsPropCode.ipynb:
  • Section 4.1: feature vector norms versus path patching scores; cosine similarities and coupling coefficients; code for getting feature scores from the Pile
  • Section 4.2: code for getting n_bias and n_subj feature scores
  • Appendix E: experiments regarding the effect of LayerNorm on feature vectors
  • Appendix G: experiments regarding the use of feature vectors for debiasing
• BestFitScores.ipynb:
  • Section 4.1: Code for calculating the best fit slope and r^2 between n_subj and n_obj feature vector scores
  • Section 4.2: Code for calculating best fit between n_subj and n_bias scores
  • Appendix I: Maximum activating examples
• OtherObservables.ipynb:
  • Section 4.3: Experiments involving other observables (political party, C vs. Python) and comparisons to other methods