The Dataset Viewer has been disabled on this dataset.

Beyond Shared Geometry

Retained coordinates, experimental results, and portable centering-analysis code for:

Beyond Shared Geometry: Predictive Recovery, Exact Actions, and Registration Preference in Language-Model Representations
Pengcheng Yang · Zenodo preprint, 2026

Paper · Full code repository · Author ORCID

This dataset release provides the original saved 16-dimensional Study 2 coordinate inputs and preserved experimental outputs. It supports analysis of fixed representation banks and the paper's template-centering diagnostic. It does not contain language-model weights or complete hidden activations.

Download format

The complete research payload is distributed as bsg_research_artifacts.zip. It contains a beyond-shared-geometry/ folder with the original directory layout, code, data, documentation, and integrity manifests. Extract the archive before running the verification or reproduction commands below. The README inside the archive preserves the earlier unpacked-package guide; for this archive-based release, follow this page's download and extraction instructions.

Download the complete archive. SHA256SUMS.txt checks the archive and the accompanying top-level documentation.

Contents after extraction

Directory inside the archive Contents
data/retained_coordinates/ 72 model/layer/domain banks with recorded readouts; coordinate NPZ files, saved PCA objects, row metadata, source manifests, and centering anchors.
results/ Preserved result tables, summaries, receipts, diagnostic outputs, and historical failure/incomplete-status evidence from the full project.
studies/template_centering_v1/ Portable input adapter, preserved scientific scripts, static verifier, and explicit numerical reproduction commands.
studies/spectral_dimension_sensitivity_v1/core/ Frozen mathematical dependency used by the centering diagnostic, with its original license notices.
upstream/ Original software citation and licensing notice, retained as provenance for the broader source project.

The coordinate export covers 24 weekday (C7), 24 square-corner (D4), and 24 supplementary S3 cells: six models, four sampled layers per model, and three domains. A cell is a model/layer/domain condition with a specified readout, not a single example. C7 and D4 use span_mean; S3 uses final_token. The export contains 147 unique original-byte objects: 72 coordinate objects, 72 PCA objects, and three shared row-metadata objects. The source manifests record model and tokenizer revisions, readout settings, array information, source archive/member hashes, and row roles.

Models represented in the export are Qwen2.5-1.5B, Qwen2.5-3B, Qwen2.5-7B, Llama-3.1-8B, Llama-3.2-1B, and Llama-3.2-3B. Their weights are not redistributed.

The coordinate objects retain multiple recorded partitions. For the C7/D4 diagnostic, the manifest binds the 128 complete test blocks per cell to the original rows. Do not treat the entire saved array as a test-only bank. The manifests, rather than filename order, define the row and object associations.

Download and inspect

This is a research-artifact repository containing NPZ, JSON, CSV, source code, and documentation. Its files are not one uniform tabular dataset, so the automatic table viewer is disabled. Download the complete snapshot with the Hugging Face CLI:

hf download yangpengchengstat/beyond-shared-geometry \
  --repo-type dataset --local-dir beyond-shared-geometry-download
cd beyond-shared-geometry-download
# On macOS/Linux, verify the downloaded archive and documentation:
shasum -a 256 -c SHA256SUMS.txt
python3 -m zipfile -e bsg_research_artifacts.zip .
cd beyond-shared-geometry

All paths below are relative to the extracted beyond-shared-geometry/ folder. Start with data/retained_coordinates/README.md and data/retained_coordinates/manifest.json. The coordinate NPZ member is z; saved PCA objects include basis and center. Use the manifest's object paths and recorded array headers to select the intended condition. Load numerical NPZ files with allow_pickle=False.

Integrity checks and reproduction

Run the package inventory check and the existing centering verifier with the Python standard library:

python3 -B verify_inventory.py
python3 -B studies/template_centering_v1/run.py verify

These commands verify files, source/result hashes, and test-row bindings. They do not extract activations, fit PCA, whiten matrices, or run the scientific analysis.

The complete C7/D4 centering workflow is included. Numerical reproduction uses NumPy 2.2.6 and is an explicit separate step:

python3 -B studies/template_centering_v1/run.py run \
  --stage rewhitened --output work/centering-rewhitened-fresh \
  --execute-offline-analysis

python3 -B studies/template_centering_v1/run.py run \
  --stage fixed --output work/centering-three-way-fresh \
  --execute-offline-analysis

The fixed-metric branch reads the authenticated saved two-treatment results, following the original workflow; it does not automatically substitute an arbitrary new first-stage output. See studies/template_centering_v1/README.md for the full protocol and dependency details.

Interpretation

The diagnostic compares the original bank, removal of the within-template state mean with the original whitening held fixed, and the same centering followed by rewhitening. The number of cells in which the prescribed image uniquely leads is:

Task Original Centered, original whitening Centered, rewhitened
Weekdays (C7) 7/24 15/24 18/24
Square corners (D4) 16/24 0/24 0/24

These are finite-bank spectral-budget rankings, not rankings of exact minimum prediction risk or evidence that a model natively computes using the prescribed action. The reported score is C16/16; in the fixed-whitening centered treatment it is not the fraction of remaining residual energy. Held-out prediction, fixed-bank constrained-risk comparisons, and spectral registration preference are distinct evaluations.

Scope and limitations

  • Full-dimensional hidden activations, model weights, extraction environments, and discarded dimensions are not included. Saved PCA bases and retained coordinates cannot recover a new PCA fitted to full-space within-template differences.
  • Separate authored-context coordinates and full historical forward caches are not included. Their saved result tables are not substitutes for missing inputs.
  • S3 coordinates/PCA/rows are included, but additional source-embedding inputs, neural fits, and interface checkpoints are not. Their reproduction requirements differ from retained-coordinate analysis.
  • Missing historical Study 1 lexical PCA outputs and operator arrays remain missing.
  • All preserved outputs retain their original interpretation and status, including failures and incomplete analyses. Inclusion does not imply every experiment succeeded.

This release enables retained-bank analysis, not end-to-end reproduction of every experiment. Some preserved README files and source manifests describe the larger upstream project and refer to files not included here. Only the centering workflow is packaged as a self-contained numerical workflow in this subset; use the full GitHub repository for other code and project-level documentation.

Provenance

Scientific code, saved coordinate objects, result files, and original manifests were copied without changing their bytes. No scientific experiment was rerun to assemble this release. The centering anchors are an already documented publication derivative of historical summaries, not original full-summary files.

UPSTREAM_FILES.json records every copied file's original path, size, SHA-256, and upstream GitHub commit. INVENTORY.json inside the archive covers the extracted scientific payload and its documentation. The top-level SHA256SUMS.txt covers this browser-upload distribution. The larger source manifests remain unchanged; they describe their original scope and are not inventories of this subset. upstream/CITATION.cff describes the upstream software, and upstream/THIRD_PARTY_NOTICES.md preserves historical licensing/status language; the current dataset license is stated below.

License and citation

The released data, saved experimental outputs, and associated data documentation are licensed under CC BY 4.0. Software, executable scripts, and software configuration snapshots retain MIT or their included original notices. See LICENSE_DATA, LICENSE_CODE, and LICENSE_SCOPE.md for the exact boundary. No license to model weights is granted by this release.

When using these artifacts, cite the paper and identify this dataset repository and the specific revision used. The following DOI identifies the paper, not a separate DOI assigned to this dataset:

@misc{yang2026beyondsharedgeometry,
  author = {Yang, Pengcheng},
  title = {Beyond Shared Geometry: Predictive Recovery, Exact Actions, and Registration Preference in Language-Model Representations},
  year = {2026},
  publisher = {Zenodo},
  note = {Preprint},
  doi = {10.5281/zenodo.24238471},
  url = {https://doi.org/10.5281/zenodo.24238471}
}

Contact: yangpengchengstat@gmail.com.

Downloads last month
43