sparse-autoencoder-training

Train sparse autoencoders by configuring LanguageModelSAERunnerConfig to decompose dense transformer activations into sparse, monosemantic interpretable features for mechanistic interpretability analysis. You can analyze polysemantic representations and superposition in models like GPT-2 and Gemma.

Sparse autoencoders are used in mechanistic interpretability to decompose dense neural activations into sparse, interpretable features. They enable researchers to analyze monosemantic representations, study superposition, and perform feature steering and ablation experiments on transformer models.

Yes, you can load pre-trained sparse autoencoders to analyze model activations across architectures such as GPT-2 and Gemma. The skill supports encoding activations to sparse features and decoding back to reconstruct activations for evaluating feature quality.

Evaluate sparse autoencoder feature quality through analysis, steering, and ablation experiments. Encode model activations to sparse features and decode them back to reconstruct activations, allowing you to assess how well the autoencoder isolates interpretable monosemantic features.

Yes, you need TransformerLens, PyTorch, and sae-lens installed. These dependencies provide the foundational framework for extracting transformer activations and configuring the sparse autoencoder training runner for mechanistic interpretability tasks.

Use sparse autoencoders when raw transformer activations are too polysemantic to interpret directly. They decompose dense activations into sparse, monosemantic features, making it easier to study superposition and isolate individual interpretable features for steering and ablation.