sida/ica-lens-paper

ICA Lens Paper Artifacts

This dataset stores public artifacts for the paper ICA Lens: Interpreting Language Models Without Training Another Dictionary.

• —Project Page: https://liusida.github.io/ica-lens-paper/
• —Code: https://github.com/liusida/ica-lens-paper
• —Interactive Demo: ICA Explorer Space

Introduction

ICA Lens is a practical workflow for stable, efficient, and auditable Independent Component Analysis (ICA) of language model representations. It recovers compact, human-interpretable directions without the need for training sparse autoencoders (SAEs). This repository contains the fitted FastICA artifacts and explorer databases for GPT-2 Small, Gemma 2 2B, and Qwen 3.5 2B Base.

Contents

models/
  gpt2/
  gemma2_2b/
  qwen3_5_2b_base/

databases/
  ica_probe_mini.sqlite
  ica_probe_full.sqlite

manifest.json
checksums.sha256

• —models/: Contains fitted FastICA artifacts and JSON metadata for the supported models.
• —databases/: Contains SQLite databases for the ICA Lens explorer. The mini database maintains core browsing functionality, while the full database preserves richer context-token data.

Sample Usage

To use these artifacts with the official ICA Lens explorer, clone the source code and run the following:

uv sync
uv run python scripts/fetch_artifacts.py --models --databases
uv run python -m server.app --port 8001

Once the server is running, the explorer will be accessible at http://127.0.0.1:8001.

Citation

@article{liu2026icalens,
  title={ICA Lens: Interpreting Language Models Without Training Another Dictionary},
  author={Liu, Sida and Han, Feijiang},
  journal={arXiv preprint arXiv:2606.11722},
  year={2026}
}