Teachafy/speakable-welfare-axes-artifacts

Jacobian lenses for Qwen3-4B-Instruct-2507 (final + penultimate targets)

Two Jacobian lenses (Gurnee et al. 2026, "workspace" paper) fitted for Qwen/Qwen3-4B-Instruct-2507 with the official jlens package, released as part of Is Functional Welfare Speakable? (Apart Digital Minds Research Sprint 2026; code: https://github.com/nsharan2000/speakable-welfare-axes).

file	target layer	prompts	dtype	notes
Qwen3-4B-Instruct-2507_jacobian_lens.pt	final (default/Neuronpedia convention)	150 wikitext	fp32	used for all headline numbers
Qwen3-4B-Instruct-2507_jacobian_lens_penult.pt	34 (penultimate; the paper's stated default for Claude models)	150 wikitext	fp32	robustness variant

Fit: jlens.fit(model, prompts, dim_batch=128, max_seq_len=128, ...), per-prompt checkpointing; all Jacobians verified finite. Per-layer top-10 lens↔model agreement for both fits: experiments/jlens-fit-2507/results/R5_lens_diagnostics.json in the repo. The penultimate-target lens agrees slightly better over the workspace band (0.171 vs 0.164 mean top-10 agreement); every result in the paper holds under both targets (j5_lens_comparison.json).

Load:

import jlens
lens = jlens.JacobianLens.load("Qwen3-4B-Instruct-2507_jacobian_lens.pt")

Positions <16 are unfitted per the jlens convention. Wikitext-fit; chat- prompt use is a stated limitation in the paper.

Also included (welfare vectors used by the paper; provenance nickmahdavi/functional-welfare, third-party reproduction of Han et al. 2026):

file	what
vectors_step95_bal.pt	trained Gold/Mold directions, RL step 95, balanced
vectors_naive_faithful_pc5000.pt	faithful-walk naive controls (5,000 traj/class)

Fetch everything into a checkout of the code repo with python3 download_artifacts.py.