sparse-autoencoder-training

You decompose language model activations into interpretable features by training Sparse Autoencoders to resolve polysemanticity. This Skill provides tools to uncover monosemantic representations, allowing you to analyze how models represent multiple concepts within single neurons.

Analyzing superposition in neural networks involves using Sparse Autoencoders to separate overlapping concepts. This Skill facilitates superposition analysis by decomposing complex activations into sparse, human-understandable features, revealing how models encode multiple representations.

You train a custom sparse autoencoder using the provided scripts to configure hyperparameters and evaluation metrics. This Skill supports training custom SAEs to discover monosemantic representations and analyze model concepts like specific entities or legal jargon.

Yes, you can use sae-lens to load pre-trained sparse autoencoders for GPT-2 small. The Skill integrates with TransformerLens, enabling you to encode model activations and discover interpretable features quickly without training from scratch.

Neural network neurons represent multiple concepts due to polysemanticity, where models superimpose many features within a single neuron. You resolve this by training Sparse Autoencoders to decompose activations into sparse, monosemantic representations for mechanistic interpretability.

Yes, this sparse autoencoder training Skill integrates with TransformerLens. It provides tools and guidance to train custom SAEs and encode model activations, allowing you to analyze superposition and discover interpretable features within language models.