br-sovereign-llm-corpus / docs /data-governance.md
guicybercode's picture
Private synthetic fixture release v0.1.0
ec0efd0 verified
|
Raw
History Blame Contribute Delete
8.13 kB

Data governance and rights protocol

Scope

This policy governs text used for tokenizer training, language-model pretraining, validation, and public dataset artifacts. It distinguishes public access from permission. A downloadable page or dataset does not enter the allowlist unless the intended copying, transformation, model training, and release are supported by document-level evidence.

This document is an engineering control, not a substitute for legal advice. Public corpus or weight release requires final legal review.

Fail-closed decision

A document is admissible only when all required conditions pass:

  1. source identity and canonical URL are known;
  2. a source revision or immutable snapshot is recorded;
  3. rights evidence is retrievable and archived by hash;
  4. commercial use and derivatives are allowed;
  5. redistribution of the released artifact is allowed;
  6. attribution and share-alike duties can be satisfied;
  7. third-party inclusions are absent or separately cleared;
  8. personal-data scanning and source-specific review pass;
  9. the document is outside takedown and quarantine lists;
  10. benchmark contamination remains below the frozen threshold.

Unknown, missing, noncommercial, or no-derivatives terms cause rejection. A human may quarantine a record for review but cannot silently override the release policy.

Initial approved source families

Portuguese Wikimedia projects

Acquire official, dated Wikimedia dumps. Text may be used under the applicable CC BY-SA terms while preserving article URL, title, revision, contributors or attribution mechanism, license URI, and change notice. Exclude images, fair-use media, and content with specific incompatible notices. Classify Brazilian and European Portuguese rather than treating every Portuguese page as pt-BR.

Rights evidence:

Portal da OBMEP

The Portal states CC BY 3.0 Brazil coverage for its material. Accept only Portal/IMPA-authored textual items whose page and embedded assets are covered by that statement. Preserve author or IMPA attribution, URL, license, and modification notice. Exclude reproduced book excerpts, external examinations, third-party figures, and material outside the Portal domain until separately cleared.

Rights evidence:

Brazilian official acts

Article 8(IV) of Law 9,610/1998 excludes official acts such as laws, decrees, regulations, and judicial decisions from copyright protection. The project initially limits this family to laws, decrees, regulations, and Section 1 of the Official Gazette. Record issuing body, act number, publication date, edition, section, page, canonical URL, and snapshot hash.

This is a statutory rights basis, not a Creative Commons license. Editorial articles, news, speeches, doctrine, visual design, and third-party attachments are not automatically covered. Gazette sections with personal data remain excluded pending LGPD review.

Evidence:

ProEdu

Evaluate every ProEdu item and file. Admit CC0, exact CC BY, and exact CC BY-SA licenses. Reject noncommercial, no-derivatives, missing-version, or missing- license records. Verify that attachments and included third-party works carry compatible rights.

Evidence:

SciELO Brasil

Admit only articles whose authoritative article record displays an exact CC BY 4.0 license. Preserve authors, title, journal, DOI, license URI, publication date, and change notice. Exclude older or ambiguous items, CC BY-NC works, and third-party figures not covered by the article license.

Rejected aggregate sources

The main corpus does not ingest these collections as wholes:

Source Reason
BrWaC academic-use restrictions and no uniform redistribution grant for underlying pages
Carolina conflicting aggregate terms and source-specific document licenses, including noncommercial terms
CulturaX package terms do not relicense mC4 and OSCAR web text
mC4/C4 dataset packaging terms do not establish rights for every source page
OSCAR corpus maintainers do not own the crawled content
HPLT release terms do not provide a homogeneous rights chain for underlying pages
Common Crawl users remain responsible for copyright, privacy, and origin-site terms
GigaVerbo and GigaVerbo-v2 mixed sources with incompatible, unknown, or restricted terms
LegalPT aggregates official acts are mixed with news, scholarship, and web-derived text
generic gov.br pages public hosting and site footers do not classify every page as an official act

A rejected aggregate can serve as a discovery index. A document must be reacquired from its original source and pass the full policy before admission.

Document contract

Every publishable document records:

  • document ID and stable group ID;
  • normalized-text and raw-content SHA-256;
  • canonical and acquisition URLs;
  • creator, publisher or issuing body, and title;
  • publication and retrieval times;
  • source revision and snapshot hash;
  • language and language variant;
  • domain and split;
  • rights basis and exact SPDX license where applicable;
  • license and evidence URLs plus evidence retrieval time and digest;
  • attribution text and share-alike status;
  • commercial-use, derivatives, and redistribution decisions;
  • third-party content status;
  • PII review state;
  • ordered transformation chain;
  • parent document ID for derived segments;
  • takedown state and review history.

Document IDs derive from stable source identity and revision, not from row position. Split assignment uses the group ID to prevent related versions from crossing evaluation boundaries.

Transformation record

Each transformation includes a name, semantic version, parameters, input hash, output hash, timestamp, and code commit. Normalization is lossless NFC plus documented control-character and whitespace handling. Material editorial changes are not hidden inside normalization.

Exact duplicate removal operates on normalized hashes. Near-duplicate removal and benchmark decontamination emit match evidence and quarantine decisions. Original raw snapshots remain immutable in the controlled artifact store when their rights permit retention.

License-separated release

Corpus shards are separated by rights family. Every record retains its own license. A mixed dataset card lists all included rights families and does not apply Apache-2.0 to source text.

Share-alike obligations are evaluated at the released corpus-artifact layer before publication. Model-weight licensing is a separate legal decision and is not inferred from the code or dataset package.

Personal data and takedown

The project minimizes personal data rather than claiming that automated PII detection is complete. High-risk sources, children's information, personal identifiers, contact details, and Gazette personnel sections are excluded from the initial corpus.

Every release has a documented removal channel and immutable takedown ledger. A removal creates a new corpus revision, records affected document and token shards, and triggers an impact review for future model releases. Historical research manifests remain available only when lawful and appropriately restricted.

Required reports

Each corpus build emits machine-readable reports for:

  • source and license decisions;
  • evidence coverage;
  • acquisition hashes;
  • encoding and language distribution;
  • PII and policy quarantines;
  • exact and near duplicates;
  • benchmark overlap;
  • split and domain distribution;
  • license-family bytes and tokens;
  • transformation lineage coverage;
  • rejected and removed records.

A later quality gate cannot downgrade an earlier legal failure into a warning.