What problem does it solve? Neuron-level units in large language models are polysemantic, making internal activations hard to interpret. This Skill decomposes a dense activation (residual stream, MLP output) into a sparse weighted sum of directions from an over-complete dictionary, turning black-box vectors into human-readable features. ## Core Features & Use Cases - Sparse Autoencoders (SAE): Train or load SAEs via SAELens to recover monosemantic features at a single activation site, with demos and references to Gemma Scope. - Transcoders: Map MLP input to output through sparse features for circuit discovery and replacement-context analysis, compatible with Anthropic's circuit-tracer. - Crosscoders: Learn a shared dictionary across layers or fine-tuned checkpoints to identify concepts introduced during chat fine-tuning. - ICA Lens baseline: Apply training-free Independent Component Analysis as a cheap first-pass interpretability lens before committing to dictionary training. - Use Case: Given a research question about what a Gemma 2 layer encodes, fit or load a dictionary, inspect top-activating features, and validate them with sparse probing or targeted perturbation. ## Quick Start Ask the assistant to decompose a chosen model layer's activations into sparse interpretable features using the SAE, transcoder, crosscoder, or ICA Lens submethod that fits your goal.
