Head Atlas

NLP Research

Overview

This project started with the goal of replicating the findings of the HeadPursuit paper and proceeded with:

  1. Extending the applicability of it’s interpretability method to any model on HuggingFace.
  2. Investigating attention head specialization metrics and their use as a stopping criterion for the SOMP1 algorithm.
  3. Visualizing the attention heads in a diagram dependent on this specialization metric.

…

The project is still ongoing.

  1. (Simultaneous Orthogonal Matching Pursuit) ↩︎