# anthropics/jacobian-lens

Companion code for the global workspace interpretability paper

Repository: https://github.com/anthropics/jacobian-lens
Canonical: https://ross.abutalabs.com/products/jacobian-lens
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-08-04T20:46:19+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 96, release rhythm 35, longevity 4
- inputs: {"age_days": 62, "days_push": 29, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1836, forks 261 (observed 2026-08-28T04:05:42.669444+00:00)

## What it is
A Python library implementing the Jacobian lens, an interpretability technique that linearly transports residual-stream activations at any layer into the final-layer basis and decodes them into ranked vocabulary tokens. It is the reference implementation accompanying the 'Verbalizable Representations Form a Global Workspace in Language Models' paper.

## Use cases
- inspect what an internal LLM activation is disposed to make the model say
- visualize layer-by-layer token readouts across positions in a prompt
- fit a Jacobian lens on any HuggingFace decoder transformer
- apply a pre-fitted lens to probe model internals
- reproduce results from the global workspace interpretability paper
- study how verbalizable representations evolve across transformer layers

## When to choose
- you want to interpret internal activations of open-weights decoder LLMs
- you need a reference implementation of the Jacobian lens technique
- you want to reproduce or extend the global workspace paper's analysis

## When to avoid
- you need a maintained, production-supported tool (it is explicitly not maintained and not accepting contributions)
- you need optimized or fast fitting (the implementation is a non-optimized reference)
- you work with non-HuggingFace model formats without adaptation

## Facets
- artifact type: library
- maturity: maintenance
- function: machine-learning, nlp, developer-tools
- domain: machine-learning, deep-learning, large-language-models
- platform: python, cross-platform
- tags: interpretability, mechanistic-interpretability, jacobian, lens, transformers, residual-stream, global-workspace, reference-implementation, research-code, natural-language-processing, research, gpu

## Member repositories
- anthropics/jacobian-lens (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:42.669444+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:18:28.236610+00:00, confidence not recorded.
  - readme: https://github.com/anthropics/jacobian-lens (fetched 2026-08-28T04:05:42.669444+00:00, sha 95970ecab791)
- Data as of 2026-08-30T08:39:29.467469+00:00.