Fitted with Anthropic's reference implementation
(
jacobian-lens, Apache-2.0)
for the paper
Verbalizable Representations Form a Global Workspace in
Language Models (2026). Averaged input-output Jacobians per layer over
wikitext-103 prompts (fit_qwen25_7b.py from that repo).
1import jlens
2lens = jlens.JacobianLens.from_pretrained(
3 "anicka/jlens-qwen2.5-7b-instruct",
4 filename="qwen2.5-7b-instruct_jlens.pt")
Companion notebook (three lenses on one activation — logit lens vs
Jacobian lens vs NLA):
https://github.com/anicka-net/nla-at-home/blob/main/notebooks/05_three_lenses.ipynb
