sparse-autoencoder-training

You can train custom Sparse Autoencoders using SAELens to decompose dense neural activations into sparse, interpretable features. The Skill guides researchers through training architectures like Standard, Gated, TopK, and JumpReLU for mechanistic interpretability tasks.

Superposition analysis examines how neural networks represent more features than dimensions by entangling them. Sparse Autoencoders help study this by decomposing dense neural activations into sparse, interpretable, monosemantic representations for clearer mechanistic insights.

Yes, SAELens works with TransformerLens to load pre-trained SAEs and run feature analysis on GPT-2 activations. This requires PyTorch-ready tooling to perform encoding, decoding, and analysis workflows on the extracted neural activations.

You can train custom Sparse Autoencoders using Standard, Gated, TopK, and JumpReLU architectures for feature discovery. These architectures decompose dense neural activations into sparse, interpretable features for mechanistic interpretability analysis.

Yes, PyTorch-ready tooling is required to load pre-trained SAEs and perform encoding, decoding, and analysis workflows. The Skill guides researchers through using SAELens and TransformerLens for these neural activation tasks.

The available analysis workflows for studying model behavior with SAEs include feature attribution, steering, and ablation. These methods help researchers study how sparse, interpretable features decompose dense neural activations and influence model behavior.