CoolFace
Datasetpublic

siddharthmb/mats-gf-provenance-corpora

Provenance-codeword training corpora All training corpora from the eight-experiment provenance codewords program (per-source activation codewords in Qwen3 models). Code, paper, and reproduction scripts: https://github.com/Sid-MB/mats-gf-provenance-codewords Each synthetic corpus ships in full: docs.parquet (training documents), train.parquet, qa.parquet (probe questions incl. phantom-fact controls), generation intermediates (raw/), the sqlite sequence store (seqdb/), and audit… See the full description on the dataset page: https://huggingface.co/datasets/siddharthmb/mats-gf-provenance-corpora.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face
0likes231downloads
Dataset Card

Provenance-codeword training corpora

All training corpora from the eight-experiment provenance codewords program (per-source activation codewords in Qwen3 models). Code, paper, and reproduction scripts: https://github.com/Sid-MB/mats-gf-provenance-codewords

Each synthetic corpus ships in full: docs.parquet (training documents), train.parquet, qa.parquet (probe questions incl. phantom-fact controls), generation intermediates (raw/), the sqlite sequence store (seqdb/), and audit reports.

Contents

FolderCorpusUsed by
v1/16 sources, planted facts (v1; answers partially guessable — the instrument flaw v2 repaired)#1
v2/repaired corpus: 16 sources x 30 facts, invented-proper-name answers (base-model guessing 0/1,920), 480 phantom questions#2, #3
e10/exact 10-exposures-per-fact subsample of v2 (per-source exact-degree MILP, seed 17)#3
capacity-256/256-source corpus, 62,720 docs, nested tier views (16/64/256), audited clean (0/6,907 base guessing)#4, #7
v3/v2 + 48 contested entities (two conflicting values, balanced claimant pairs, exposure cells 40:40/40:20/40:10), phantom-contested + off-corpus controls#8
natural-rebuild/rebuild bundle for the natural CC-News corpus (see below)#9

The natural corpus is a rebuild recipe, not raw text

The #9 corpus weaves 480 invented tracer facts into 12,480 real news articles from 16 outlets (CC-News). Raw CC-News article text is not redistributed here (news-article licensing). Instead, natural-rebuild/ contains everything else — outlet census and selection, per-article manifest (URL, normalized-text SHA-1, dedup cluster, role/split, doc id), tracer facts, sentence banks, QA, mined-fact tables, and per-document content hashes — plus a deterministic rebuild script:

bash
pip install datasets pandas pyarrow
python natural-rebuild/scripts/rebuild_natural_corpus.py \
    --bundle-dir natural-rebuild --out-dir rebuilt/

It streams the public vblagoje/cc_news dataset, matches each article by (URL, normalized-text SHA-1), re-runs the seeded weaving step, and verifies every rebuilt document against the recorded content_sha1 — the result is bit-identical to the corpus the experiment trained on. --dry-run checks the first few matches in ~minutes.

Provenance and license

Built 2026-06/07. Source experiment IDs and cluster paths: ARTIFACTS.md in the GitHub repo. Synthetic corpora and manifests are CC-BY-4.0. All planted facts are invented; phantom-fact tables mark the guessing floors. Companion repos: adapters, demo checkpoint, readouts.