What problem does it solve? Training sparse autoencoders or transcoders to find interpretable directions in a language model is expensive and slow. This Skill provides a training-free alternative: it applies Independent Component Analysis (FastICA) to cached activations at any hook point, recovering candidate monosemantic directions without fitting a dictionary, so researchers can inventory, score, and compare interpretable structure quickly. ## Core Features & Use Cases - Activation capture and ICA fitting: Stream activations from residual stream, MLP output, or attention-head hook points and fit FastICA via config-driven workflows. - Component scoring and comparison: Rank directions by non-Gaussianity, run SAEBench targeted probe perturbation and sparse probing, and measure overlap against trained SAE features. - Explorer UI: Browse, annotate, and share fitted components through a FastAPI-based explorer backed by SQLite databases, with pre-fitted artifacts for GPT-2 Small, Gemma 2 2B, and Qwen 3.5 2B Base. - Use Case: Before committing GPU budget to SAE training on a new model layer, capture activations, fit ICA, rank directions by non-Gaussianity, and inspect the top candidates in the explorer to decide whether a learned dictionary is even necessary. ## Quick Start Ask the agent to clone the ica-lens-paper repository, run uv sync, fetch the released artifacts, and launch the explorer server on port 8001 to browse pre-fitted ICA components.
