Pre-fitted Jacobian lenses computed through bnb 4-bit quantized networks, for reading and steering 4-bit local deployments.

These are distinct from neuronpedia/jacobian-lens, which were fitted on bf16. These lenses were fitted on 4-bit models, which should help match inference closer.

Each model folder holds lens.pt, its meta.json, and a README with the source checkpoint, revision, quantization, and fit parameters.