lamm-mit/gemma4-interpretability

Gemma materials-science interpretability research archive

Research records supporting Reading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Model, Markus J. Buehler. Release identifier: paper-revision-2026-09-06.

This archive supplies the original observations, supporting state arrays, exact prompts, protocols, intervention records, statistics, analysis source, and generated research figures. It includes the original 4B readout and geometry studies; steering and patching; supplied-equation and equation-free relational experiments; 12B/31B scale extensions; and the three-scale coordinate intervention and retrospective Comment 3 overlap audit. Development and diagnostic records needed to understand those experiments are retained, including negative results and superseded controls. Not every archived development record is a confirmatory endpoint in the paper.

The archive preserves the original jlens_materials/ layout. MANIFEST.json and FILE_INVENTORY.csv enumerate the research-record files with their sizes, SHA-256 hashes, and inclusion basis. Release documentation and verification reports are separately identified; they were prepared after the experiments. RESULT_MAP.csv maps manuscript figures and tables to supporting records and reproduction commands. The final companion manuscript sources, figures, compact SI data, and clean/highlighted PDFs are under `paper/`; paper/SOURCE_MANIFEST.json identifies that package separately from the original data snapshot.

Download and verify

Use the full dataset commit hash cited by the revised paper/release index:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="lamm-mit/gemma4-interpretability", repo_type="dataset",
                  revision="e5f5080bdd9ed99ef99b8f81b31f9e93f2de4075", local_dir="gemma4-interpretability")

The hash above retrieves the immutable original research-data snapshot cited in the paper. main additionally contains the final companion manuscript package and updated release documentation; the original research records are unchanged. From the downloaded directory run python verify_release.py. The checker verifies every original research-record file against the manifest. REPRODUCIBILITY.md describes analysis-only reproduction and its validation scope. Work in a copy when rerunning historical analysis scripts because they write derived outputs in place.

Protocol chronology and corrections

“Frozen,” “prespecified,” and historical filenames containing “preregistration” describe the authors' internal research records. They do not, by themselves, establish independently timestamped preregistration. The public GitHub repository was created on 22 July 2026, after the July experiments. This data release is later still. CHRONOLOGY.json and CHRONOLOGY.csv preserve recorded dates and status statements without treating editable timestamps or hashes as external timing evidence. Analyses developed after inspection of data remain post hoc even when a protocol preceded the subsequent calculation.

Original files are retained byte for byte, including historical terminology and original validator labels. For the causal-coordinate experiment, the original validator checked distinct law identifiers; a later audit found three repeated physical relations in the development and target cohorts. The corrected interpretation and sensitivity results are in jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/ and revised SI S7A/Table S45. Historical classifications in original files should be read alongside that correction. Superseded Gaussian control diagnostics are retained for provenance and excluded from the amended primary inference.

Related resources

• —Public source repository
• —Fitted Jacobian lenses: three 4B, three 12B, and one 31B paper fits.
• —Original exploratory latent vectors
• —Reusable prompt dataset

The Gemma base checkpoints and fitted lens weights are not duplicated here. They are unnecessary for recomputing analyses from the saved observations and states; rerunning model inference requires the exact checkpoint/lens revisions and applicable access conditions. The included Apache-2.0 LICENSE and NOTICE originate from the public source repository and cover its source code; upstream model and third-party resource terms remain applicable.