Transcoder-circuits: reverse-engineering LLM circuits with transcoders

This repository contains tools for understanding what's going on inside large language models by using a tool called "transcoders". Transcoders decompose MLP sublayers in transformer models into a sparse linear combination of interpretable features. By using transcoders, we can reverse-engineer fine-grained circuits of features within the model.

To get started, we recommend working through the walkthrough.ipynb notebook. The full structure of the repository is as follows:

• Installation tools:
  • setup.sh: A shell script for installing dependencies and downloading transcoder weights.
  • requirements.txt: The standard Python dependencies list.
• Examples:
  • walkthrough.ipynb: A walkthrough notebook that demonstrates how to use the tools provided in this repository for reverse-engineering LLM circuits with transcoders.
  • train_transcoder.py: An example script for training a transcoder.
• Experimental results;
  • sweep.ipynb: An evaluation of transcoders and SAEs trained on Pythia-410M (weights not provided).
  • interp-comparison.ipynb: Code for performing a blind comparison of SAE and transcoder feature interpretability
• Case studies:
  • case_study_citations.ipynb: An example of a reverse-engineering case study that we carried out, in which we investigated a transcoder feature that activates on semicolons in parenthetical citations.
  • case_study_caught.ipynb: An example of a reverse-engineering case study that we carried out, in which we investigated a transcoder feature that activates on the verb "caught".
  • case_study_local_context.ipynb: An example of a reverse-engineering case study that we carried out, in which we attempted to reverse-engineer a circuit that computes a harder-to-interpret transcoder feature. (We were less successful in this case study, but are including it in the interest of transparency.)
  • restricted blind case studies.ipynb: A notebook containing a set of "restricted blind case studies" that reverse-engineer random GPT2-small transcoder features without looking at MLP0 transcoder feature activations.
• Libraries:
  • sae_training/: Code for training and using transcoders. The code is largely based on an older version of Joseph Bloom's excellent SAE repository -- shoutouts to him!. (The misnomer sae_training is a vestige of this origin of the code.)
  • transcoder_circuits/: Code for reverse-engineering and analyzing circuits with transcoders. These are the tools that we use in the walkthrough notebook and in the case studies.