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 2048-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 (2048-D vectors) · reduce (PCA / t-SNE / UMAP + HDBSCAN clusters).

Tổng quan · At a glance

Chỉ số · Metric	Giá trị · Value
Văn bản · Documents	1,963
Câu · Sentences (across the corpus)	273,379
Có cấu trúc · With structure layer	1,963
Có số án lệ · With precedent number	51
Có embedding · With embedding vector	1,963
Có projection · With reduce projections	1,963
Trung vị trang · Median pages / doc	9
Trung vị ký tự · Median chars / doc	19,939
Trung vị đoạn văn · Median paragraphs / doc	47
Trung vị câu · Median sentences / doc	122

Phân loại · Document classes

Loại văn bản · doc_type

Value	Count	Share
ban_an	1,155	58.8%
quyet_dinh	784	39.9%
unknown	24	1.2%

Lĩnh vực · case_type

Value	Count	Share
dan_su	860	43.8%
hinh_su	481	24.5%
hanh_chinh	308	15.7%
unknown	163	8.3%
kinh_doanh_thuong_mai	99	5.0%
hon_nhan_gia_dinh	44	2.2%
lao_dong	8	0.4%

Cấp xét xử · doc_subtype

Value	Count	Share
phuc_tham	1,104	56.2%
giam_doc_tham	718	36.6%
unknown	105	5.3%
so_tham	18	0.9%
tai_tham	18	0.9%

Cấp toà · court_level

Value	Count	Share
cap_cao	1,731	88.2%
unknown	138	7.0%
toi_cao	70	3.6%
tinh	13	0.7%
huyen	11	0.6%

Lược đồ bảng documents · documents schema

The default config carries one row per document with three families of columns:

Identification + meta

Field	Type	Description
doc_name	string	Stable document id (== source dDocName query parameter).
source	string	Source host, always anle.toaan.gov.vn.
detail_url / pdf_url	string	Deep link back to the portal page / PDF.
doc_code	string	E.g. 38/2021/DS-PT (sequence/year/case-type-procedure).
doc_type	string	ban_an | quyet_dinh | an_le | ban_cao_trang.
case_type	string	dan_su | hinh_su | hon_nhan_gia_dinh | lao_dong | kinh_doanh_thuong_mai | hanh_chinh.
doc_subtype	string	so_tham | phuc_tham | giam_doc_tham | tai_tham | an_le.
year	int32	Year extracted from doc_code.
title	string	Header line as captured.
subject	string	V/v ... matter line.
issue_date	string	ISO 8601 issue date when discoverable.
issuing_body	string	Full court name.
court_level	string	huyen | tinh | cap_cao | toi_cao.
jurisdiction	string	Province / city qualifier extracted from the body.

Body + stats

Field	Type	Description
markdown	string	NFC-normalised, modern-orthography Vietnamese markdown (page-segmented with ## Page N headings).
num_pages / num_sections / num_paragraphs / num_sentences	int32	Counts from the structure layer.
char_len	int32	Character length of markdown.
text_hash	string	SHA-256 first-32 hex of markdown (re-run-stable id).
parser_model / parsed_at	string	Provenance for the parse stage.

Hierarchy + entities

Field	Type	Description
structure_json	string	Full DocumentStructure (meta + stats + sections + paragraphs + sentences) as JSON; round-trips via json.loads.
extracted_json	string	Regex NER + statute-link output (entities, relations, statute_refs) as JSON.

Precedent layer (án-lệ-only)

Field	Type	Description
precedent_number	string	E.g. Án lệ số 47/2021/AL. None for plain judgments.
adopted_date	string	ISO 8601 adoption date.
applied_article_code / applied_article_number / applied_article_clause	string / int64 / int64	Most-cited statute reference.
principle_text	string	"Nội dung án lệ" / "Nguyên tắc" excerpt when present.

Quick load:

import json
from datasets import load_dataset

ds = load_dataset("tmquan/anle-toaan-gov-vn", split="train")  # documents config
row = ds[0]
structure = json.loads(row["structure_json"])
print(structure["meta"]["doc_code"])
for sec in structure["sections"]:
    print(sec["kind"], sec["label"])

Companion stages · sentences + embed + reduce

Alongside the default documents-*.parquet shards (one row per document, with markdown + structure), the dataset ships up to three additional parquet bundles that mirror the extract → embed → reduce stages 1-to-1. All three join back to the documents table on the doc_name primary key.

sentences-*.parquet — sentence-level rows

One row per sentence; the full hierarchical structure (section → paragraph → sentence) is exposed as flat parquet rows so consumers can stream, filter, and embed sentences directly without parsing the structure_json blob.

Field	Type	Description
doc_name	string	Join key back to documents-*.parquet.
sentence_id	string	Stable per-corpus sentence id (<doc_id>::s<g> form).
paragraph_id / section_id	string	Parent paragraph / section ids.
case_type / doc_type / doc_subtype / court_level / year / precedent_number	string / int32	Parent-document filter columns (promoted so consumers can slice without joining).
section_kind	string	header | case_summary | findings | decision | footer.
paragraph_kind	string	text | numbered_finding | numbered_decision | list_item | signature | ...
paragraph_marker	string	The marker as it appears in the body (e.g. [1], 1., -).
page	int32	Page number inside the parent PDF.
index_in_paragraph / global_index	int32	Position inside the parent paragraph / inside the document.
char_start / char_end	int32	Char span back into the parent markdown.
text	string	The sentence itself, NFC-normalised.

Quick load:

from datasets import load_dataset


sents = load_dataset("tmquan/anle-toaan-gov-vn", "sentences", split="train")
print(sents[0]["text"])
# Filter to all sentences from civil-law cassation precedents
civil_cassation = sents.filter(
    lambda r: r["case_type"] == "dan_su" and r["doc_subtype"] == "giam_doc_tham"
)

embed-*.parquet — dense vectors

One row per embeddable document (1,963 rows). The default embedder is nvidia/llama-nemotron-embed-1b-v2 (2048-D). The full set of predefined models the pipeline can route to is published in manifest.json["pipeline"]["embed"]["registry"]:

Model	Dim	Native window	Notes
nvidia/llama-nemotron-embed-1b-v2	2048	8192	Default. 1B params, 8k context.
nvidia/llama-3.2-nv-embedqa-1b-v2	1024	512	Previous ViLA default (retrieval, 512-tok window).
nvidia/llama-embed-nemotron-8b	4096	8192	8B params, higher quality.
sentence-transformers/paraphrase-multilingual-mpnet-base-v2	768	128	Multilingual MPNet (50+ langs incl. VI).
microsoft/harrier-oss-v1-270m	—	32768	270 M params, 32k native context.
microsoft/harrier-oss-v1-0.6b	—	32768	Lightweight HF default; 32k native.
microsoft/harrier-oss-v1-27b	—	32768	Highest quality in the harrier-oss family.

Field	Type	Description
doc_name	string	Join key back to documents-*.parquet.
text_hash	string	SHA-256 of the post-normalisation markdown.
embedding	list<float32>	2048-D dense vector.
embedding_dim	int64	Length of embedding (denormalised for fast filtering).
embedding_model_id	string	Model slug as the backend reports it.
embedding_text_hash	string	SHA-256 of the exact text fed to the embedder.
embedding_chunks_used	int64	Windows mean-pooled into the final vector (1 if the doc fits in one window).
embedding_chunking	string	off / sliding / sentence.

reduce-*.parquet — 2D projections + cluster ids

One row per embeddable document. Every method in ['pca', 'tsne', 'umap'] runs over the full embedding matrix with n_components=2; the clusterer is hdbscan (label -1 = noise bucket).

Field	Type	Description
doc_name / text_hash	string	Join keys back to documents-*.parquet and embed-*.parquet.
pca_x	float64	PCA projection (axis x).
pca_y	float64	PCA projection (axis y).
tsne_x	float64	TSNE projection (axis x).
tsne_y	float64	TSNE projection (axis y).
umap_x	float64	UMAP projection (axis x).
umap_y	float64	UMAP projection (axis y).
cluster_id	int64	HDBSCAN cluster label; -1 is the noise bucket.

Join back to documents:

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 embedding · Embedding visualization

Mỗi điểm là một văn bản; toạ độ là vector embedding 2048-D từ nvidia/llama-nemotron-embed-1b-v2 chiếu xuống 2D bằng UMAP, cụm bằng HDBSCAN. Mỗi hình một hàng. — Each dot is one document; coordinates are the 2D UMAP projection of a 2048-D embedding from nvidia/llama-nemotron-embed-1b-v2, with HDBSCAN cluster ids. One figure per row.

PCA và t-SNE vẫn được tính sẵn và lưu trong reduce-*.parquet (pca_{x,y,z}, tsne_{x,y,z}) — chỉ không vẽ trong card này. — PCA and t-SNE coordinates are still pre-computed and shipped in reduce-*.parquet (pca_{x,y,z}, tsne_{x,y,z}); they are simply not rendered inline here.

UMAP colored by case_type

UMAP colored by doc_subtype

UMAP colored by court_level

UMAP colored by cluster_id

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/llama-nemotron-embed-1b-v2 (2048-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: 2026-05-21T08:04:01.313084+00:00.

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. Bản phân phối lại này dùng giấy phép CC-BY-4.0; 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. This redistribution is shared under CC-BY-4.0; 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), 2048-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).}
}