Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherence - Source Code

This repository provides the source code and experimental flows required to reproduce and follow the monosemanticity analysis presented in the paper "Measuring Monosemanticity in Sparse Autoencoders via Latent Activation Coherence".

The paper was accepted for the ECML PKDD 2026 workshop.

In this Zenodo repository you can find:

• environment configuration descriptions (in two forms: as a Python's Requirements.txt file and a .devcontainer for fully separated environments);
• experiments pipeline as a set of consecutive scripts, presenting the operation in a step-by-step manner;
• source code for all the required data fetching, calculations and analysis.

With the use of this repository you may conduct and observe the Tversky Monosemanticity Score (TMS) application as a label-free metric for monosemanticity operationalization.