dacheah/bbnj-high-seas-treaty-corpus
BBNJ / High Seas Treaty Corpus A neutral, provenance-first, machine-readable record of the 2023 BBNJ Agreement (the "High Seas Treaty", in force 17 January 2026) in all six authentic UN languages, and its implementing framework. Every record carries its official source, retrieval date, citation, authentic language, an authoritative-status flag, a SHA-256 content hash, and an honest per-language fidelity flag (extracted_verified → extracted_unverified → ocr_unverified). Source… See the full description on the dataset page: https://huggingface.co/datasets/dacheah/bbnj-high-seas-treaty-corpus.
BBNJ / High Seas Treaty Corpus
A neutral, provenance-first, machine-readable record of the 2023 BBNJ Agreement (the "High Seas Treaty", in force 17 January 2026) in all six authentic UN languages, and its implementing framework. Every record carries its official source, retrieval date, citation, authentic language, an authoritative-status flag, a SHA-256 content hash, and an honest per-language fidelity flag (extracted_verified → extracted_unverified → ocr_unverified).
- Source of truth / build history: https://github.com/dacheah/bbnj-high-seas-treaty-corpus
- Human-browsable site: https://dacheah.github.io/bbnj-high-seas-treaty-corpus/ (enable GitHub Pages)
- 15 instruments · 16 document versions (16 verified against official sources) · 924 provisions · 2906 neutral concept tags
Why this dataset is different
Most legal datasets are scraped text with no sourcing. Here every row is traceable: the authoritative text is stored in its authentic language with a content hash and a fidelity flag, and anything generated (concept tags, structure) is kept in a separate, clearly-labelled layer. That makes it low-risk input for retrieval-augmented generation and citation-grounded legal NLP.
Configs
`documents` — one row per instrument version: the full authoritative text plus provenance (source_url, official_citation, authoritative_status, text_fidelity, content_hash, license, rights_note, provenance_note, dates, jurisdiction, language).
`provisions` — one row per structural unit (article / section / principle / guideline): the unit text, its neutral concepts (a model pass, unreviewed), and a link back to the source instrument's text hash and URL. Ideal as pre-chunked, concept-labelled context for RAG.
from datasets import load_dataset
docs = load_dataset("dacheah/bbnj-high-seas-treaty-corpus", "documents")
prov = load_dataset("dacheah/bbnj-high-seas-treaty-corpus", "provisions")Contents
Licensing
Mixed, and recorded per record. The compilation, structure, and concept tags (the derived layer) are CC BY 4.0. Source texts are not relicensed — each keeps its own terms (e.g. UN-materials terms; public-domain government works), recorded in every row's license / rights_note. See the licensing policy.
Disclaimer
This is a reference record, not legal advice. The authoritative text of each instrument is the original in its authentic language; translations and concept tags are unofficial and derived (the concept tags are an unreviewed model pass — expect occasional over/under-tagging). Always confirm against the cited official source for anything legally operative.
Citation
BBNJ / High Seas Treaty Corpus (Dan Cheah, @dacheah). https://github.com/dacheah/bbnj-high-seas-treaty-corpusDataset generated from the repository by `scripts/exporthfdataset.py` on 2026-08-02 — do not edit by hand.