Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders

TL;DR

This paper introduces a scalable approach to extract millions of features from Llama-3.1-8B using Sparse Autoencoders, enabling detailed interpretability and analysis of language models.

Contribution

The authors develop and evaluate 256 Sparse Autoencoders trained on all layers of Llama-3.1-8B, including modifications to Top-K SAEs, and provide open-source tools and checkpoints.

Findings

SAEs generalize to longer contexts and fine-tuned models

Feature splitting reveals new features in learned representations

Publicly available SAE checkpoints facilitate interpretability research

Abstract

Sparse Autoencoders (SAEs) have emerged as a powerful unsupervised method for extracting sparse representations from language models, yet scalable training remains a significant challenge. We introduce a suite of 256 SAEs, trained on each layer and sublayer of the Llama-3.1-8B-Base model, with 32K and 128K features. Modifications to a state-of-the-art SAE variant, Top-K SAEs, are evaluated across multiple dimensions. In particular, we assess the generalizability of SAEs trained on base models to longer contexts and fine-tuned models. Additionally, we analyze the geometry of learned SAE latents, confirming that \emph{feature splitting} enables the discovery of new features. The Llama Scope SAE checkpoints are publicly available at~\url{https://huggingface.co/fnlp/Llama-Scope}, alongside our scalable training, interpretation, and visualization tools at \url{https://github.com/OpenMOSS/Language-Model-SAEs}. These contributions aim to advance the open-source Sparse Autoencoder ecosystem and support mechanistic interpretability research by reducing the need for redundant SAE training.