sparse-autoencoder-training

To train sparse autoencoders for feature discovery, you can use SAELens to optimize sparsity and reconstruction on transformer activations. This decomposes neural activations into interpretable feature representations for mechanistic interpretability research.

Mechanistic interpretability analyzes language model internals, and sparse autoencoders help by decomposing complex neural activations into sparse, monosemantic features. This reveals specific representations within transformer activations.

Yes, you can load pre-trained sparse autoencoders with SAELens to inspect feature activations per token. This allows immediate analysis of sparse feature representations on sample prompts without training from scratch.

Yes, SAELens integrates with TransformerLens for end-to-end experimentation. You can load a model, attach a pre-trained sparse autoencoder, and analyze sparse feature activations directly within the TransformerLens workflow.

You need SAELens, TransformerLens, and PyTorch installed, plus a compatible language model and dataset setup. This environment enables training, analysis, and evaluation of sparse autoencoders for feature discovery.

You can apply feature steering and ablation to study model behavior and safety-relevant representations using sparse autoencoders. SAELens enables targeted manipulation of identified interpretable features within transformer activations.