Skip to content

Latest commit

 

History

History
23 lines (13 loc) · 1.53 KB

README.md

File metadata and controls

23 lines (13 loc) · 1.53 KB
Raw

Swap graphs: Discovering the role of neural network components at scale

This librairy is an implemetation of input swap graphs described in this post. It is a tool to uncover the role of neural networks component by using causal interventions.

The librairy is built on top of TransformerLens. The code base is in very early stage and under active development. Feel free to contact me at alexandre.variengien@gmail.com if you have questions about the code or wants to make serious use of it.

I'd recommand starting with this colab demo.

For more advanced example that uses swap graphs to craft validation experiments, you can explore the tests directory.

Install

pip install git+https://github.com/aVariengien/swap-graphs.git

Scripts

We also provide scripts to handle swap graphs at scale.

  • compute_sgraphs.py is used to compute a swap graphs for every component at a given position (often the last position in a sequence).
  • plot_semantic_maps.py uses the fiels created by compute_sgraphs.py to create the semantic maps visualisation.
  • sgraph_causal_scrubbing.py runs causal scrubbing experiments where all components up to layer L are scrubbed.
  • targetted_rewrite.py (only for the IOI dataset) runs targetted rewrite experiments for the senders and extended name mover heads.