README.md

metadata

pretty_name: BR Sovereign LLM Corpus
language:
  - pt
license: apache-2.0
task_categories:
  - text-generation
tags:
  - brazilian-portuguese
  - pretraining
  - provenance
  - licensed-data
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/documents.parquet

BR Sovereign LLM Corpus

Status

This public repository is an audited corpus protocol and initial validation snapshot. Public release 0.1.1 contains only a small, explicitly identified corpus sample for pipeline verification. It is not a target-scale pretraining corpus and does not support model-quality claims.

The scientific corpus remains under construction. Aggregate counts, source shares, and token counts are not reported until a content-addressed snapshot passes the legal, privacy, lineage, duplicate, and contamination gates.

Materialized validation snapshot

The underlying snapshot contains eight synthetic pt-BR documents drafted with AI assistance for this project, reviewed and adopted for release by the project owner. All eight records passed the production license and provenance boundary; none was rejected, quarantined, or marked as a duplicate. This verifies the data path but is not scientific training data.

Artifact	SHA-256
source JSONL	6896439e772f212c2848846f1f908410a606d81cd8845583024c16496a07979f
corpus manifest file	afbd872b05f61ffc8acf11e436949fb760f9456a3eda62dc11a67ed1511d8841
logical corpus manifest	fcb1af582f5ded8ecbe2b67e1da2b391f3fcf17678feba6389cf4a211d82ff5d
audit report file	c99b951ab4cf79356930e3ea5fdfe8269aaa6cbc7142fa6c26a054ce32f02d9e
logical audit report	af0dbeb4f1f75e3dce14bc5f317b584bcd4b7a662e48c2365f11efe484ed0711
Parquet file	e2b5d4c01739e92240cfe99e6dccebce3903729d083e61e86afa37f6dd86c8a2

The snapshot is published with its Parquet file, source JSONL, source and license registries, audit report, and content-addressed manifest.

Dataset purpose

The dataset supports research on from-scratch Brazilian Portuguese language-model pretraining under document-level rights and provenance controls. Every publishable record must link its normalized text to:

• a canonical source and source revision;
• a raw snapshot hash;
• a rights basis and exact license when applicable;
• a digest of the rights assertion and, for scientific records, the archived evidence bytes and media type;
• attribution and share-alike requirements;
• an ordered transformation chain;
• language variant, domain, stable group, and split;
• PII, third-party-content, and takedown decisions.

Public availability alone is not sufficient. Missing, unknown, noncommercial, and no-derivatives rights fail closed.

Initial source policy

Eligible source families are:

• dated official Portuguese Wikimedia dumps under applicable CC BY-SA terms;
• verified Portal da OBMEP items under CC BY 3.0 Brazil;
• Brazilian official acts covered by Article 8(IV) of Law 9,610/1998;
• ProEdu items with exact CC0, CC BY, or CC BY-SA terms;
• SciELO Brasil articles with exact item-level CC BY 4.0 terms.

The initial release rejects aggregate ingestion from BrWaC, Carolina, CulturaX, mC4, OSCAR, HPLT, Common Crawl, GigaVerbo, GigaVerbo-v2, and mixed LegalPT collections. Those packages do not by themselves establish a homogeneous rights chain for the underlying text.

Published fixture schema

The Parquet file in this release contains exactly these columns:

document_id, text, normalized_content_sha256, split, normalization_id, source_id, source_url, canonical_url, retrieved_at, license_id, rights_basis, attribution_text, creator, publisher, publication_date, source_revision_id, source_snapshot_sha256, contains_third_party, pii_status, provenance_sha256, and evidence.

Planned scientific-corpus contract

The future scientific corpus is planned to include the following core fields. This table is a contract for a later release, not a description of additional columns hidden in the current fixture:

Field	Meaning
document_id	stable source and revision identifier
group_id	split-safe work or revision family
text	normalized training text
source_id	audited source-registry identifier
source_url	acquired record URL
canonical_url	authoritative public identity
source_revision_id	immutable source revision
source_snapshot_sha256	acquired snapshot digest
raw_content_sha256	raw item digest
normalized_content_sha256	normalized text digest
language	BCP 47 language tag
language_variant	Brazilian, European, mixed, or undetermined
domain	registered content domain
rights_basis	license, statutory exclusion, or other reviewed basis
license_spdx	exact SPDX expression where applicable
license_url	authoritative license text
evidence_url	rights evidence
license_evidence_sha256	digest of the recorded evidence assertion
attribution_text	required attribution
transformation_chain	ordered versioned transformations
split	deterministic group-aware split
pii_status	personal-data review state
takedown_status	active, quarantined, or removed

Additional fields are documented in the research repository. A record missing a required release field is rejected.

Licensing

Public release 0.1.1 contains only the eight project-controlled synthetic fixtures. The owner releases them under Apache-2.0 to the extent of his rights; AI-assisted text may not be unique. No third-party corpus record is present.

A future mixed scientific corpus will use source-specific terms and must not be added under this blanket metadata. Wikimedia, Creative Commons, and statutory source records would retain their applicable terms and attribution. Model-weight licensing remains a separate decision.

Processing

The registered flow is:

1. immutable source acquisition and hashing;
2. rights-evidence validation;
3. text extraction;
4. lossless NFC and control-character normalization;
5. language, quality, PII, and policy classification;
6. exact and near duplicate detection;
7. benchmark decontamination;
8. deterministic group-aware splitting;
9. license-separated Parquet release;
10. content-addressed token-shard construction.

Every transformation records code version, parameters, input digest, and output digest.

Limitations

• The current sample is too small and too synthetic for language modeling.
• Registry statement hashes are tamper-evident metadata, not independent legal proof. Scientific records additionally require hashes of archived source or license bytes before inclusion.
• Automated PII, language, and duplicate checks are imperfect.
• A source-family allowlist does not replace item-level review.
• Share-alike effects on any future combined release require final legal review.
• Takedowns create new corpus revisions and may affect future model releases.
• The current eight-document fixture has no external benchmark content and no observed personal data; this does not validate future corpus content.

Maintenance

Initial author and maintainer:

Guilherme Monteiro, FIAP
ORCID: https://orcid.org/0009-0008-5294-224X

For correction, attribution, or withdrawal requests, open a Community discussion in this Hugging Face repository and cite the commit, document ID, and requested action. Do not post sensitive personal data; request a private follow-up when necessary. A credible urgent report can cause the repository to be hidden during review. Confirmed changes receive a new version and manifest.

The scoped approval and rights review are published in PUBLICATION_REVIEW.md; the machine-readable removal history is takedown-ledger.json.