sparse-autoencoder-training

Train sparse autoencoders using SAELens to decompose neural activations into sparse, interpretable features. Configure SAE architectures like Standard, Gated, or TopK with specific expansion factors to study monosemantic representations in language models.

Feature discovery in mechanistic interpretability identifies interpretable features within neural networks. You need sparse autoencoders when analyzing superposition or studying monosemantic representations across transformer activations to understand internal model representations.

Yes, you can load pretrained SAEs via SAELens to examine feature activations on sample GPT-2 layers. This allows you to begin interpreting sparse representations and perform feature attribution without training from scratch.

Supported sparse autoencoder architectures include Standard, Gated, and TopK models with configurable expansion factors. These architectures help analyze superposition and assess monosemantic representations in transformer language models.

Apply trained SAEs to decompose transformer activations into sparse features, enabling feature attribution and feature steering. The workflow includes training, analysis, and steering phases with documented references for understanding model representations.

Sparse autoencoders resolve superposition by decomposing dense transformer activations into interpretable, monosemantic features. Direct analysis struggles with polysemantic neurons, whereas SAEs provide a structured approach to feature discovery and mechanistic interpretability.