j-lens

Minimal implementation of the Jacobian lens and J-space methods from Verbalizable Representations Form a Global Workspace in Language Models (Gurnee et al., 2026), on Qwen3.5-4B.

• jlens.py — fitting the lens ($J_\ell = \mathbb{E}[\partial h_{\text{final}} / \partial h_\ell]$), readout, lens vectors, sparse J-space decomposition
• interventions.py — steering, ablation, and lens-coordinate swaps
• evals.py — the §A.6 pass@k evaluations and §4.1 workspace-band metrics
• main.ipynb — the full run: fit/load → read out → evaluate → intervene

Eval prompt data is downloaded from the official companion repo (anthropics/jacobian-lens) on first use.

pip install torch transformers datasets huggingface_hub matplotlib