sparse-autoencoder-training

SAELens and TransformerLens are required to load transformer models and train sparse autoencoders. These frameworks provide the necessary workflows for activation analysis, feature discovery, and evaluating reconstruction quality.

Sparse autoencoder feature discovery decomposes dense neural network activations into sparse, interpretable representations. This allows researchers to identify specific concepts represented within language model activations for mechanistic interpretability studies.

Yes, you can load and analyze pretrained Sparse Autoencoders to perform feature attribution, steering, and ablation. This supports mechanistic interpretability experiments by inspecting learned features within transformer model activations.

Evaluating sparse autoencoder reconstruction quality involves measuring controlled sparsity and reconstruction metrics during training. SAELens workflows provide the necessary tools to assess how accurately the autoencoder decomposes transformer activations into interpretable features.

Yes, this skill supports activation steering, ablation, and superposition studies on transformer models. It applies SAELens workflows to inspect learned features and perform mechanistic interpretability experiments on neural network activations.