Jacobian Lens Playground
Chat with a model and inspect the J-Space of its latest response.
Qwen/Qwen3.6-27B · 27B
Send a message to see what appears in the model’s J-Space.
Free Chat shows only exact checkpoints with a validated assistant template. Base-only lenses, including the currently published Gemma 4 adapters, are hidden because substituting an -it checkpoint would invalidate the fitted lens.
Apply the intervention to your next message, or regenerate the last turn. Use one vocabulary token per field; copy ␣word from the readout to preserve a leading space.
Send a message to populate the J-Space.
Select a cell to inspect candidates across the complete conversation.
Select a result cell after sending a message.
Choose a comparison model in Settings.
The ranked readout aggregates retained candidates over only the latest assistant turn, including its generation-prefix positions. Chat-template controls, EOS markers, and whitespace-only candidates are excluded from the ranked concepts. The blue strip shows the layers where each token appears.
JACOBIAN transports activations through the fitted average downstream Jacobian. LOGIT uses the model unembedding directly. DIFF compares them.
STEER adds a normalized token direction at every selected layer. SWAP exchanges two J-lens coordinates with the paper's pseudoinverse operation, leaving the residual component outside their span unchanged.
Lens fitting is a batch GPU job, not an interactive ZeroGPU request. The
repository includes scripts/fit_and_publish.py, which packs a text corpus,
checkpoints the fit so it can resume, writes the compatible catalog metadata,
and uploads the result to a Hugging Face model repository. About 100 prompts
is useful for a first lens; the released lenses use 1,000.
pip install -r requirements-fit.txt
python scripts/fit_and_publish.py \
--model Qwen/Qwen3-4B \
--repo your-name/jacobian-lenses \
--key qwen3-4b-community --prompts 100
Set JLENS_EXTRA_REPOS=your-name/jacobian-lenses on the Space to add every
compatible lens in that repository to the model catalog.
Method and code by Anthropic · fitted lenses by Neuronpedia.