ICA Lens

ICA Lens interprets language-model activations with Independent Component Analysis. It is substantially more compute-efficient to fit than an SAE dictionary and supports base and instruction-tuned language models.

Documentation · 中文文档 · Paper · Model collection

Get started

pip install icalens

Load a published Lens and analyze text:

from icalens import ICALens

lens = ICALens.from_pretrained("sida/icalens-gpt2-small-pile10k")
result = lens.analyze("She deposited the check at the bank.", layer=6)
result

Browse the ICA Lens model collection for other published lenses.

In Jupyter or Colab, the final result expression displays an interactive token-level analysis:

Use signed ICA scores or switch the explorer to per-token component energy. Save the same view as a standalone HTML file with:

result.to_html("analysis.html")

The first analysis loads the language model and requested Lens layer. Later calls on the same lens reuse the model in memory. device="auto" uses CUDA when available and otherwise uses the CPU.

Analyze conversations

Instruction-tuned models accept completed conversations using the standard {role, content} format:

lens = ICALens.from_pretrained("sida/icalens-qwen3.5-2b-ultrachat-1m")
result = lens.analyze(
    [
        {"role": "user", "content": "What is the most interesting science?"},
        {"role": "assistant", "content": "Physics."},
    ],
    layer=16,
)
result

Chat templates are applied automatically, and template tokens and message turns are grouped in the interactive result.

Steering

Generate normally or clamp a signed ICA coordinate during generation:

messages = [{
    "role": "user",
    "content": "If you had to pick one, what is the most interesting science? Be brief.",
}]

baseline = lens.generate(messages, max_new_tokens=16)
steered = lens.generate(
    messages,
    layer=5,
    clamp=(188, -20.0),
    max_new_tokens=16,
)

Component labels, signs, and suitable targets must be established empirically for the exact Lens and layer. See the steering tutorial for the reproducible inspection and calibration workflow.

Fit a Lens

Run a small GPT-2/Pile-10k example with the installed CLI:

icalens fit text \
  --model openai-community/gpt2 \
  --dataset NeelNanda/pile-10k \
  --layers 6 \
  --token-budget 1000 \
  --max-iter 20 \
  --output icalens-output/gpt2-demo

Fit an instruction-tuned model from UltraChat conversations:

icalens fit chat \
  --model Qwen/Qwen3.5-2B \
  --dataset HuggingFaceH4/ultrachat_200k \
  --layers 12 \
  --token-budget 100000 \
  --output icalens-output/qwen-demo

ICA Lens includes a PyTorch FastICA implementation and does not depend on SciPy or scikit-learn. Blockwise fitting and layer-at-a-time capture support larger token collections while bounding memory use.

Profile every fitted layer

After fitting, profile the components against a representative corpus:

icalens profile \
  --lens icalens-output/gpt2-demo \
  --layers all \
  --dataset NeelNanda/pile-10k \
  --split train \
  --max-tokens 10000

Profiles add sign statistics, high-energy examples, Logit Lens tokens, and optional R-lens readouts to the existing Lens directory. They help label and inspect components without changing the fitted directions.

Publish to Hugging Face

Authenticate with hf auth login, set HF_TOKEN, or add a .env file in the current directory containing a write-enabled token:

HF_TOKEN=hf_...

Then publish the saved Lens as a Hugging Face model repository:

icalens publish \
  --lens icalens-output/gpt2-demo \
  username/icalens-gpt2-demo

The artifact records the analyzed model, activation site, fitted layers, preprocessing, component profiles, and fitting and profiling provenance. Individual layer and profile files are downloaded lazily when a published Lens is used.

Learn more

The documentation covers:

• Getting started
• Text and chat
• Component profiles
• Scores and energy
• Steering
• Reconstruction
• Fitting and publishing
• Python API

Authors

• Sida Liu
• Feijiang Han

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}
}