ATTRIBUTION
This repository is CC BY 4.0 without qualification. It contains only what this work produced. Copyright (c) 2026 Tarush Gupta.
What is here
| what | size | licence |
|---|---|---|
results/**/*.pt: 380 fitted Jacobian maps and 2 activation caches |
16.2 GB | CC BY 4.0 |
results/**/*.json, results/**/*.npz: experimental records, plus 222 map sidecars |
28.5 MB | CC BY 4.0 |
results/found_shard0.tgz: E35's per-worker containment bitmaps for shard 0 of the Pythia training stream; results/r7_box_sha256*.txt: E74's on-box hash lists |
13 MB | CC BY 4.0 |
corpora/manifest.json, corpora/manifest_ood.json, corpora/dataset_revisions.json |
13 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 corpus plaintext is not distributed: eleven files built from seven public datasets, plus
one derived from them. It is 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 with the code repository, github.com/tarush-ai/fittingcorpus:
bash repro/06_data.sh --build # builds at the pinned revisions, then hash-verifies against the manifests
bash repro/06_data.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.
OOD_News_2024_dedup has no upstream of its own: the manifest records the rule that derives it
from OOD_News_2024.
corpora/dataset_revisions.json pins the upstream dataset commit for all eight upstreams the code
reads: the seven corpus sources, and the Pythia training stream the containment scores are measured
against. Both builders (src/build_corpora.py, src/build_ood_corpora.py) read it and refuse to load a
dataset that has no pinned revision.
Measured 2026-08-31, rebuilding from scratch on a clean machine (the day before the revision pins were added, so this is also evidence that the upstreams had not moved by then):
| 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) |
OOD_News_2024 and its derived _dedup |
not in that record |
Every corpus in the last two rows carries its SHA-256 in the manifest, and --build checks it.
A pinned revision protects against an upstream edit, but not against an upstream deletion. If
a source dataset or its pinned commit is removed from the Hub, that corpus cannot be rebuilt; the
manifest hash still says which file the results were computed on.