This repository hosts Transluce's circuit tracing codebase, as described in the blogpost.

First clone this repo:

Then install the dependencies using uv:

You must set environment variables in order to use some parts of the codebase. Copy the template file and fill in the missing values:

The main codebase is in the circuits directory. The lib directory contains shared libraries used by the codebase, which have been also used in prior releases by Transluce (e.g. observatory).

The following subdirectories live in circuits:

• clso: Core implementation of gradient-based node attribution and edge weight tracing in our work. Also includes helpers for applying RelP and Integrated Gradients on arbitrary LMs.
• evals: Evaluation code for the SVA benchmark, adapted from Marks et al. (2024). The evaluations in our blogpost can be replicated with this code.
• analysis: Codebase for analysing and visualising circuits, taking in inputs from the algorithms implemented in clso. Includes utilities for feature scoring given hypotheses, clustering, and steering. All data is stored and processed as pandas dataframes.

Additionally, scripts includes utilities for replicating our evaluations and case studies.

For an example of how to trace and save circuits over a dataset, see the scripts/case_studies/math/math.py script.

If you use this codebase in your research, please cite our blogpost: