tmquan/anle-toaan-gov-vn

Vietnamese Án lệ Corpus — anle.toaan.gov.vn

🇻🇳 Tóm tắt. Bộ dữ liệu đa cấp của các bản án + án lệ Việt Nam thu thập từ cổng anle.toaan.gov.vn của Tòa án nhân dân tối cao. Mỗi văn bản pháp lý đi kèm markdown đã chuẩn hoá tiếng Việt (NFC + chính tả hiện đại) và lớp cấu trúc phân cấp đầy đủ: document → section → paragraph → sentence. Bộ dữ liệu ship bốn cấu hình HF tương ứng bốn giai đoạn pipeline: documents (mặc định) · sentences (mức câu) · embed (vector 4096-D) · reduce (PCA / t-SNE / UMAP + cụm HDBSCAN).

🇬🇧 Summary. Multi-level corpus of Vietnamese court judgments and precedents (án lệ) harvested from anle.toaan.gov.vn (Supreme People's Court portal). Every document carries a Vietnamese-normalised markdown body (NFC + modern orthography) and a hierarchical structure layer: document → section → paragraph → sentence, every unit carrying a stable id + char span back into the markdown. The dataset ships four HF configurations matching the four pipeline stages: documents (default) · sentences (sentence-level rows) · embed (4096-D vectors) · reduce (PCA / t-SNE / UMAP + HDBSCAN clusters).

Tổng quan · At a glance

Chỉ số · Metric	Giá trị · Value
Văn bản · Documents	1,844
Câu · Sentences	169,698
Trích dẫn luật · Law citations (sentence-grounded)	29,783
Trích dẫn bản án · Case citations	4,238
Án lệ chính thức · Official án lệ (is_precedent)	9
Có embedding · Embedding vectors (4096-d)	1,844
Có projection · Reduce projections (PCA/t-SNE/UMAP)	1,844
Trung vị trang · Median pages / doc	9
Trung vị ký tự · Median chars / doc	20,909
Trung vị câu · Median sentences / doc	79

Phân loại · Document classes

Column names and categorical values are English (for adoption + education); legal content — court, law_name, citation ref text, code, markdown — stays Vietnamese.

Lĩnh vực · category

Value	Count	Share
Civil	892	48.4%
Criminal	459	24.9%
Administrative	326	17.7%
Commercial	105	5.7%
Marriage & Family	47	2.5%
Labor	8	0.4%
null	7	0.4%

Cấp xét xử · instance_level

Value	Count	Share
Appellate	1,014	55.0%
Cassation (giám đốc thẩm)	680	36.9%
First-instance	126	6.8%
Retrial (tái thẩm)	17	0.9%
null	7	0.4%

Cấp toà · court_level — parsed from the issuing-court letterhead

Read from the level keyword right after Tòa án nhân dân (tolerant of two-column letterheads and anonymised place names); district markers take precedence over a following province marker. Fill 98.9%.

Value	Count	Share
High (cấp cao)	1,721	93.3%
Supreme (tối cao)	70	3.8%
Provincial (tỉnh)	20	1.1%
District (huyện/quận)	12	0.7%
null	21	1.1%

(High dominates because án lệ sources are overwhelmingly cấp-cao appellate/cassation rulings.)

Lược đồ bảng documents · documents schema

One row per judgment; primary key doc_name (all configs join on it). 28 columns:

Identity — doc_name (TAND portal id, PK) · source · web_url (portal page) · pdf_url (PDF binary).

Official document id — official_document_id (printed, e.g. 237/2022/HS-PT) · official_document_id_normalized (hyphenated, for joins/citation-graph) · number · year · code (VI ký hiệu) · id_source (regex | audit_corrected).

Classification — category · instance_level · court (VI court name) · court_level (English tier) · issued_date (date) · date_source.

Precedent — precedent_number (e.g. 44/2021/AL, non-null on the 9 official án lệ) · is_precedent (bool).

Citations — first-class list<struct>:

Column	Struct fields
citations_law (29,783)	kind (provision|document) · ref (VI, ordered large→small Chương→Mục→Điều→Khoản→Điểm + law + năm) · chapter,section,article,clause,point · law_type,law_name (VI) · id,year · sentence_id (grounding → sentences) · span (char offsets [[start,end],…] into markdown)
citations_case (4,238)	id,number,year,code · role (first-instance|appellate|cassation|protest) · domain,level · sentence_id,span

Plus citations_source (llm | llm+section | regex), num_law_citations, num_case_citations. 79% of law citations carry a sentence_id + span (anchored on Điều <article>; the rest are normative-by-number ids not found verbatim).

Content — markdown (full judgment) · markdown_chars · num_pages · confidence · flags (list).

Quick load:

from datasets import load_dataset

docs = load_dataset("tmquan/anle-toaan-gov-vn", split="train")   # documents config
r = docs[0]
print(r["official_document_id"], r["category"], r["instance_level"], r["court"])
for c in r["citations_law"][:5]:
    print(c["ref"], "→", list(c["sentence_id"]))   # citation + its grounding sentence(s)

Companion configs · sentences + embed + reduce

Three parquet bundles accompany documents; all join on the doc_name primary key, and sentences also links to citations via sentence_id.

sentences-*.parquet — sentence-level rows (169,698)

Sentence-split from the shipped markdown (conservative Vietnamese-legal splitter); char_start/char_end index the same markdown, so citations_*.span and citations_*.sentence_id line up exactly.

Field	Type	Description
sentence_id	string	{doc_name}#{sent_idx} — the grounding id referenced by citations_*.sentence_id.
doc_name	string	Join key back to documents.
sent_idx	int	Sentence index within the document.
char_start / char_end	int	Char span into the parent documents.markdown.
text	string	The sentence.
category / instance_level / year / precedent_number	string / int	Parent-document facets (promoted for slicing without a join).

from datasets import load_dataset
sents = load_dataset("tmquan/anle-toaan-gov-vn", "sentences", split="train")
# every sentence that grounds a citation of Điều 468
hits = sents.filter(lambda r: "Điều 468" in r["text"])

embed-*.parquet — dense vectors (1,844)

nvidia/Nemotron-3-Embed-8B-BF16 (4096-d, "Nemotron-3" = v3), chunking=sentence (sentence-boundary windows + mean-pool + L2-norm) with the model's passage: input prefix, served via vLLM (native Ministral kernel) on a GB10.

Field	Type	Description
doc_name	string	Join key.
embedding	list<float32>	4096-d dense vector (L2-normalised).
embedding_dim	int	4096.
embedding_model_id	string	nvidia/Nemotron-3-Embed-8B-BF16.

reduce-*.parquet — 2D projections (1,844)

PCA + t-SNE + UMAP fit over the full embedding matrix with n_components=2 (HDBSCAN clustering is off in this release).

Field	Type	Description
doc_name	string	Join key.
pca_x / pca_y	float	PCA 2D projection.
tsne_x / tsne_y	float	t-SNE 2D projection.
umap_x / umap_y	float	UMAP 2D projection.

from datasets import load_dataset
docs   = load_dataset("tmquan/anle-toaan-gov-vn", "documents", split="train").to_pandas()
embed  = load_dataset("tmquan/anle-toaan-gov-vn", "embed",     split="train").to_pandas()
reduce = load_dataset("tmquan/anle-toaan-gov-vn", "reduce",    split="train").to_pandas()
joined = docs.merge(embed, on="doc_name").merge(reduce, on="doc_name")

Trực quan hoá · Visualizations

Sơ đồ Sankey trích dẫn · Citation Sankey — category · subcategory → document → cited provision

Mỗi luồng là một trích dẫn pháp luật: bắt đầu từ lĩnh vực · cấp xét xử của bản án, đi qua tài liệu (toàn bộ 1.844 bản án), tới điều khoản được viện dẫn (Điều → Khoản → Điểm, kể cả Chương/Mục). Màu luồng theo lĩnh vực·cấp. — Each flow is one law-citation: from the judgment's legal category · court level, through the document (all 1,844 judgments), to the cited provision (Điều → Khoản → Điểm, incl. Chương/Mục). Links are coloured by category·subcategory.

Bản tương tác (zoom/hover) · Interactive version: sankey-category-document-citation.html

Chiếu embedding 2D · 2D embedding projections

Mỗi điểm là một văn bản; toạ độ là vector embedding 4096-D từ nvidia/Nemotron-3-Embed-8B-BF16 chiếu xuống 2D bằng t-SNE và UMAP. Toạ độ PCA/t-SNE/UMAP đều lưu trong reduce-*.parquet. — Each dot is one document; coordinates are the 2D t-SNE and UMAP projections of a 4096-D nvidia/Nemotron-3-Embed-8B-BF16 embedding. PCA/t-SNE/UMAP coordinates all ship in reduce-*.parquet.

PCA — theo lĩnh vực · by legal category

t-SNE — theo lĩnh vực · by legal category

UMAP — theo lĩnh vực · by legal category

Cách thu thập + chuẩn hoá · How the corpus was built

The pipeline is a five-stage NeMo Curator flow (download → parse → extract → embed → reduce) defined under packages/datasites/anle:

1. Download — walks the paginated Nguồn án lệ + curated Án lệ listings, downloads each PDF.
2. Parse — pypdf for digital PDFs, falls back to nvidia/nemoretriever-parse for image-only scans (hybrid runtime). Output is NFC-normalised Vietnamese markdown with modern orthography.
3. Extract — three deterministic layers:
   • Generic — regex + dictionary NER (dates, courts, articles, precedent numbers) and statute linking (Điều N khoản M Bộ luật ...).
   • Site (precedent) — normalises án lệ metadata onto a stable schema (precedent_number, adopted_date, applied_article_*, principle_text).
   • Structure — segments markdown into the canonical five-section template (header → case_summary → findings → decision → footer), paragraphs (with marker classification: numbered_finding [1], numbered_decision 1., list_item -, text, signature), and sentences (regex split on [.?!] + capital).
4. Embed — default model: nvidia/Nemotron-3-Embed-8B-BF16 (4096-D vector). Sliding-window chunking + mean-pool when a doc exceeds the model's native context window. The set of predefined embedding models the pipeline can route to is published in manifest.json["pipeline"]["embed"]["registry"].
5. Reduce — every method in pca, tsne, umap runs over the full embedding matrix; HDBSCAN labels the cluster id. cuML on a GPU worker; sklearn / umap-learn / hdbscan otherwise.

All five layers are deterministic and re-runnable.

Captured: 2026-07-21T16:46:56.048226+00:00.

Hạn chế & lưu ý · Limitations & caveats

• Tone-mark erratum (repaired 2026-07). Parquet revisions published before July 2026 carried a normalizer bug that moved tone marks inside closed syllables (hoạt → họat, toàn → tòan, ngoài → ngòai), affecting 1,962 of 1,963 documents. The text columns in this revision are repaired (the swap is length-preserving, so char_len and entity offsets stay valid). Two consequences remain: text_hash / embedding_text_hash still hash the pre-repair text (kept as stable join keys), and the shipped embeddings were computed from the pre-repair sentences — the drift is confined to tone-mark placement within affected syllables.
• jurisdiction is deprecated. See the schema note: it is a lossy, fully re-derivable substring of issuing_authority, records the court's seat rather than its territorial jurisdiction, and is null for ~65% of rows. It will be dropped in the next schema revision.
• Weak / regex labels. doc_type, case_type, doc_subtype, court_level and the extraction layers are deterministic regex heuristics over OCR-imperfect PDF text, not human annotation. Expect residual noise in all classification columns, and OCR noise (letterhead dashes, truncated names) in issuing_authority.
• Court-taxonomy vintage. court_level follows the pre-2025 court organization (huyen/tinh/cap_cao/toi_cao). Vietnam's 2025 court reorganization replaced district-level courts with regional courts (Tòa án nhân dân khu vực); documents in this corpus predate that change.

Dữ liệu cá nhân · Personal data & anonymization

Bản án, quyết định trên cổng của Toà án nhân dân tối cao được mã hoá thông tin cá nhân trước khi công bố theo Nghị quyết 03/2017/NQ-HĐTP của Hội đồng Thẩm phán: tên đương sự được thay bằng ký hiệu (ví dụ "Nguyễn Văn A", "bà B"), địa chỉ và số giấy tờ được rút gọn. — Judgments and decisions on the Supreme People's Court portal are anonymized by the courts before publication under Resolution 03/2017/NQ-HĐTP of the Council of Judges: party names are replaced with coded forms (e.g. "Nguyễn Văn A", "Mrs. B"), addresses and ID numbers are truncated.

• Names of judges, prosecutors and court clerks (public officials acting in their official capacity), case numbers and dates remain in the text as published.
• Anonymization is applied by the issuing court and can be incomplete in places (e.g. organization names, land-parcel or vehicle identifiers). This dataset redistributes the portal text verbatim — no additional anonymization and no de-anonymization was performed.
• Do not attempt to re-identify anonymized parties. Downstream processing of personal data must comply with Vietnam's personal-data framework (Decree 13/2023/NĐ-CP and the Personal Data Protection Law effective 2026) and any law applicable to you (e.g. GDPR).

Nguồn · Source

• Portal: https://anle.toaan.gov.vn/
• Publisher: Supreme People's Court of Vietnam (Tòa án nhân dân tối cao)

Giấy phép · License

Văn bản gốc được Toà án nhân dân tối cao công bố trên cổng thông tin công cộng. Theo Điều 15 Luật Sở hữu trí tuệ, văn bản tư pháp không thuộc phạm vi bảo hộ quyền tác giả, nên giấy phép CC-BY-4.0 ở đây chỉ áp dụng cho phần tuyển chọn, cấu trúc hoá, chú giải và embedding do bộ dữ liệu này bổ sung — không áp dụng cho văn bản toà án gốc. Vui lòng kiểm tra điều khoản sử dụng của trang nguồn trước khi tái phân phối thương mại. — The source documents are published by the Supreme People's Court on a public portal. Under Article 15 of Vietnam's Intellectual Property Law, judicial documents are excluded from copyright, so the CC-BY-4.0 grant here covers only the curation, structuring, annotations and embeddings added by this distribution — not the underlying court texts. Please check the source-website terms of use before commercial redistribution.

Trích dẫn · Citation

If you use this dataset, please cite both the redistribution on Hugging Face and the original source (Tòa án nhân dân tối cao):

@misc{anle_2026,
  title        = {Vietnamese Án lệ + Bản án Corpus (anle.toaan.gov.vn)},
  author       = {TMQuan},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/tmquan/anle-toaan-gov-vn}},
  note         = {Multi-level mirror with hierarchical structure (DocumentMeta + Section + Paragraph + Sentence), 4096-D embeddings, and 2D projections over the Vietnamese án-lệ portal.}
}

@misc{anle_toaan_2026,
  title        = {Vietnamese Án lệ + Bản án Corpus},
  author       = {{Án lệ — Tòa án nhân dân tối cao}},
  year         = {2026},
  howpublished = {\url{https://anle.toaan.gov.vn/}},
  note         = {Official portal for Vietnamese án lệ (precedents) + nguồn án lệ (precedent source materials), published by the Supreme People's Court (Tòa án nhân dân tối cao).}
}