sparse-autoencoder-training

To train a sparse autoencoder, you use SAELens and TransformerLens to decompose dense model activations into a sparse latent space. This workflow provides configuration options and training routines to discover interpretable features within your transformer models.

Mechanistic interpretability uses sparse autoencoders to study superposition and concept learning by breaking down dense activations into interpretable features. You need sparse autoencoders when trying to understand what specific concepts a transformer model has learned internally.

Yes, analyzing sparse autoencoders with SAELens requires PyTorch, TransformerLens, and SAELens as specified dependencies. These tools provide the necessary framework to load models, encode activations, and inspect the resulting sparse features.

Feature-based steering or ablation is performed by training a sparse autoencoder to expand activations into sparse features, then manipulating those specific features. This workflow enables you to isolate interpretable features and alter them to steer model behavior.

Sparse autoencoder architectures are compared using evaluation metrics provided within the training workflow to guide end-to-end experiments. These metrics help assess how effectively the autoencoders decompose activations and produce interpretable features.

Yes, you can reproduce experiments by loading a pre-trained SAE and a transformer model to encode activations and inspect sparse features. The workflow includes tutorials and reference materials to compare SAE architectures and validate findings.