sparse-autoencoder-training

Train sparse autoencoders for neural network interpretability using this framework to decompose model activations into interpretable features. It provides scripts and guidance leveraging SAE Lens and TransformerLens to analyze activation geometry and feature discovery.

Superposition in language models is a phenomenon where networks represent more features than dimensions. You can analyze it by training sparse autoencoders on model activations to identify and isolate these interpretable concepts using the provided analysis tools.

Identify safety-relevant features like deception by applying sparse autoencoder analysis to model activations. This Skill helps you decompose activations into interpretable features, enabling the detection of harmful content and safety analysis within neural networks.

Yes, you need TransformerLens and PyTorch to run sparse autoencoder training. The Skill requires Python with specific dependencies including sae-lens, transformer-lens, and torch to properly execute the training and feature analysis scripts.

Yes, you can use sparse autoencoders to understand feature geometry across different neural network architectures. The framework supports training SAEs on various models to analyze activation structures and discover interpretable features.

The best way to start discovering interpretable features is by installing sae-lens and following the provided tutorials. This Skill offers a framework to train SAEs and analyze the resulting features to understand model behavior and activation geometry.