sparse-autoencoder-training

Use this Skill to train sparse autoencoders on transformer activations with sparsity controls and checkpointing. It decomposes neural network activations into interpretable features while tracking reconstruction metrics and sparsity regularization throughout the training process.

Feature steering with sparse autoencoders attributes model outputs to specific SAE features and steers generation along feature directions. You can test causal importance by zeroing selected features to observe how specific interpretable concepts affect downstream predictions.

Yes, you can load existing pretrained sparse autoencoders and inspect which features activate on specific prompts or tokens. The Skill evaluates pretrained or custom SAEs through SAELens, summarizing top activating features and reconstruction quality for transformer models.

Analyze superposition by training sparse autoencoders to decompose transformer activations into sparse, interpretable features. This reveals how models represent overlapping concepts within neural network layers, allowing you to identify syntax or named entities and measure their downstream effects.

Yes, TransformerLens is required for precise handling of hook activations during sparse autoencoder training and analysis. The Skill relies on TransformerLens to extract transformer model activations that are then decomposed into interpretable features via SAELens.

Limitations of sparse autoencoders for feature discovery include reliance on sparsity regularization quality and reconstruction metrics. Interpretable features require careful configuration of sparsity controls, and causal claims about feature importance depend on ablation results rather than activation patterns alone.