ATTRIBUTION
This repository is CC BY 4.0 without qualification. It contains only what this work produced.
What is here
| what | size | licence |
|---|---|---|
results/lenses/** — 382 fitted operators + sidecars |
16.2 GB | CC BY 4.0 |
results/**.json, results/**.npz — experimental records |
~27 MB | CC BY 4.0 |
corpora/manifest.json, corpora/manifest_ood.json, corpora/dataset_revisions.json |
12 KB | CC BY 4.0 |
Derived from Pythia (EleutherAI, Apache-2.0) via the Jacobian-lens reference
implementation (Apache-2.0, pinned 581d398). Neither restricts these derivatives.
What is deliberately NOT here, and how to get it
The eleven corpus plaintext files are not distributed. They are third-party text under
upstream terms; Wikipedia_en.jsonl and OOD_Wikipedia_2023.jsonl are Wikipedia-derived and
carry CC BY-SA, which is share-alike and cannot be relicensed as CC BY 4.0. Shipping them
would have forced a mixed-licence repository and made the CC BY 4.0 claim above false.
Rebuild the pool from its public sources:
bash repro/06_build_corpora.sh # builds, then hash-verifies against the manifests
bash repro/06_build_corpora.sh --verify # verify a pool you already have
corpora/manifest.json and corpora/manifest_ood.json ship here and carry, per corpus: the
SHA-256 of the exact file used, the document count, the source dataset id and config, and the
three disjoint seed-block index ranges. A rebuild that matches those hashes is the same pool.
Measured 2026-08-31, rebuilding from scratch on a clean machine:
| corpus | rebuild |
|---|---|
| Pile-CC, StackExchange, Wikipedia_en, Github, USPTO_Backgrounds | 5/5 bit-identical |
| OOD_CommonPile, TRAP_FineWeb, CONTROL_PubMed_2023 | 3/3 bit-identical |
corpora/manifest.json regenerated |
byte-identical |
| OOD_Wikipedia_2023, OOD_arXiv_2023 | not re-run (53 MB / 110 MB of streaming); both carry SHA-256 in the manifest and are checked by the script |
One caveat, stated plainly. src/build_corpora.py calls load_dataset( "monology/pile-uncopyrighted", ...) with no revision= pin, so an upstream change to that
dataset would produce a different pool. The result above is evidence the upstream had not moved
as of that date — not a guarantee it never will. The manifest is the check: a mismatch means the
upstream moved, which is a finding to report, not a failure to work around. The OOD panel is more
stable by construction because its configs are date-stamped (20231101.en, 2024-06).