sparse-autoencoder-training

To train a sparse autoencoder for language model activations, configure sparsity, architecture, and logging parameters to study superposition and monosemanticity. This workflow supports custom SAE training integrated directly with TransformerLens.

Monosemanticity analysis in neural network interpretability identifies sparse, interpretable features within dense model activations. By training sparse autoencoders, researchers isolate individual features to study how specific concepts are represented.

Yes, you can load pretrained SAEs with TransformerLens and Neuronpedia. This capability allows you to load pretrained sparse autoencoders for common model layers, inspect features on real prompts, and browse them using Neuronpedia.

SAELens supports feature steering and ablation workflows through activation encoding and decoding. You can analyze feature attribution, ablate specific features, and test how steering influences model predictions and behavior.

The best way to analyze superposition in language models is by training custom sparse autoencoders. Transforming dense activations into sparse features reveals how models represent more concepts than available neurons by overlapping representations.