--- language: - vi license: cc-by-4.0 pretty_name: "Vietnamese National Legal Database (vbpl.vn)" size_categories: - 100K 🇻🇳 **Tóm tắt.** Bộ dữ liệu mức **văn bản** của > **Cơ sở dữ liệu Quốc gia về pháp luật** thu thập từ cổng > [`vbpl.vn`](https://vbpl.vn/) do Bộ Tư pháp vận hành. Bao gồm > luật, pháp lệnh, nghị định, thông tư, quyết định, nghị quyết, > chỉ thị… ở cả cấp **trung ương** (Quốc hội, Chính phủ, các bộ) > lẫn **địa phương** (HĐND/UBND 63 tỉnh, thành). Mỗi dòng là một > văn bản pháp luật kèm markdown đã chuẩn hoá tiếng Việt > (NFC + chính tả hiện đại sau 1984) và lớp cấu trúc phân cấp đầy > đủ: **document → section → paragraph → sentence**. > > 🇬🇧 **Summary.** Document-level corpus of Vietnam's > **National Legal Database** harvested from > [`vbpl.vn`](https://vbpl.vn/) (operated by the Ministry of Justice). > Covers laws, ordinances, decrees, circulars, decisions, > resolutions, directives, etc. across both **central** (National > Assembly, Government, ministries) and **provincial** (People's > Council / People's Committee of the 63 provinces and cities) > levels of authority. Each row is one legal document with its > Vietnamese-normalised markdown body (NFC + modern post-1984 > orthography) and a hierarchical structure layer: > **document → section → paragraph → sentence**, every unit > carrying a stable id + char span back into the markdown. ## Tổng quan · At a glance | Chỉ số · Metric | Giá trị · Value | |---|---:| | Văn bản công bố · Documents | **158,822** | | Tổng số bản ghi đầu vào · Raw extract rows | 158,822 | | Loại bỏ vì rỗng · Dropped (empty body) | 0 | | Không có thân bài · `markdown` is null (source has no body) | 11,505 | | Có cấu trúc · With structure layer | 158,822 | | Có tệp đính kèm · With downloaded attachment | 0 | | Trung vị ký tự · Median chars / doc | 5,111 | | Trung vị trang · Median pages / doc | – | | Trung vị đoạn văn · Median paragraphs / doc | 1 | | Trung vị câu · Median sentences / doc | 17 | ## Tổng quan trực quan · Visual overview Sáu hình tóm tắt cấu trúc của corpus. Tất cả hình được tạo bằng `plotly` + `kaleido` (Chromium headless) và lưu thành PNG tĩnh để hiển thị trong dataset card. — Six pictures summarising the corpus structure. All figures are rendered with `plotly` + `kaleido` (headless Chromium) and saved as static PNG for the dataset card. ### Phân bố theo lĩnh vực pháp luật ![Phân bố theo lĩnh vực pháp luật](./overview-legalarea-treemap.png) Top 25 legal areas sized by document count. The pale `Chưa phân loại` ('uncategorised') rectangle covers ~72% on its own -- vbpl.vn editors haven't tagged most legacy docs. ### Phạm vi → loại văn bản ![Phạm vi → loại văn bản](./overview-scope-doctype-sunburst.png) Two-level radial split: `scope` (inner) → `doc_type` (outer, top 12 per scope). Reveals where the corpus weight sits -- `dia_phuong` (~66%) is mostly `QĐ` (decisions) and `NQ` (resolutions), while `trung_uong` mixes `QĐ` / `NĐ` (decrees) / `TT` (circulars). ### Top 20 loại văn bản ![Top 20 loại văn bản](./overview-doctype-bars.png) Document types ranked by count, with the canonical short code, full Vietnamese name, and an English gloss. ### Số văn bản theo năm và phạm vi ![Số văn bản theo năm và phạm vi](./overview-year-stack.png) Stacked area of documents issued per year, split by `scope`. The legacy CMS migration concentrates early years; modern post-2010 output is dominated by `dia_phuong`. ### Loại văn bản qua các năm ![Loại văn bản qua các năm](./overview-doctype-year-heatmap.png) Top-12 `doc_type` × year heatmap (log₁₀ scale). Shows the shift from `CT` directives in the early 90s to `QĐ` decisions as the workhorse instrument from 2000 onwards. ### Top 15 cơ quan ban hành ![Top 15 cơ quan ban hành](./overview-agency-bars.png) Top issuing agencies. Quốc hội + Chính phủ + the larger ministries dominate `trung_uong`; provincial People's Councils / Committees split the `dia_phuong` half. ## Phạm vi · Scope split Bộ dữ liệu chia làm hai nhánh: ``trung_uong`` (văn bản pháp luật do Quốc hội + Chính phủ + các bộ ngành Trung ương ban hành) và ``dia_phuong`` (HĐND/UBND của 63 tỉnh, thành). — The corpus splits into ``trung_uong`` (central authorities: National Assembly, Government, ministries) and ``dia_phuong`` (the 63 provinces and cities, mostly People's Council / People's Committee output). | Value | Count | Share | |---|---:|---:| | `dia_phuong` | 104,345 | 65.7% | | `trung_uong` | 54,477 | 34.3% | ## Loại văn bản · `doc_type` + `legal_type` Mỗi văn bản được gắn **slug tiếng Việt không dấu** (`doc_type`, ví dụ `quyet_dinh`, `nghi_dinh`, `thong_tu_lien_tich`) lẫn tên đầy đủ tiếng Việt (`legal_type`, ví dụ `Quyết định`, `Nghị định`, `Thông tư liên tịch`). Slug được sinh tự động từ tên đầy đủ qua [`packages.datasites.vbpl.codes.slugify_vi`](https://github.com/tmquan/ViLA/blob/main/packages/datasites/vbpl/codes.py) nên người đọc hiểu ngay loại văn bản mà không cần tra cứu bảng mã. Mã ngắn cũ (`QĐ`, `NĐ`, `TTLT`, ...) vẫn xuất hiện trong số hiệu (`43/2026/NĐ-CP`) và có thể khôi phục từ `SLUG_TO_CANONICAL_CODE`. — Each document carries a self-describing **ASCII snake_case slug** (`doc_type`, e.g. `quyet_dinh`, `nghi_dinh`, `thong_tu_lien_tich`) and the canonical Vietnamese full name (`legal_type`, e.g. `Quyết định`, `Nghị định`, `Thông tư liên tịch`). The slug is automatically derived from the full Vietnamese name via [`slugify_vi`](https://github.com/tmquan/ViLA/blob/main/packages/datasites/vbpl/codes.py) so a reader can interpret the doc type without consulting a separate codebook. The compact short code (`QĐ`, `NĐ`, `TTLT`, ...) still appears inside `so_hieu` itself (`43/2026/NĐ-CP`) and can be recovered from any slug via the `SLUG_TO_CANONICAL_CODE` table in `codes.py`. The set of slugs follows Luật Ban hành Văn bản Quy phạm Pháp luật 2015 — `hien_phap` / `bo_luat` / `luat` / `phap_lenh` / `lenh` / `nghi_quyet` / `nghi_dinh` / `quyet_dinh` / `thong_tu` / `chi_thi` / `sac_lenh` / `van_ban_hop_nhat` / `thong_tu_lien_tich` / ... ### By slug · `doc_type` | Value | Count | Share | |---|---:|---:| | `quyet_dinh` | 84,970 | 53.5% | | `nghi_quyet` | 26,091 | 16.4% | | `thong_tu` | 16,212 | 10.2% | | `ban_dich_van_ban` | 10,576 | 6.7% | | `chi_thi` | 7,908 | 5.0% | | `nghi_dinh` | 4,690 | 3.0% | | `thong_tu_lien_tich` | 3,393 | 2.1% | | `van_ban_hop_nhat` | 1,792 | 1.1% | | `sac_lenh` | 980 | 0.6% | | `cong_van` | 708 | 0.4% | | `luat` | 580 | 0.4% | | `lenh` | 376 | 0.2% | ### By full name · `legal_type` | Value | Count | Share | |---|---:|---:| | `Quyết định` | 84,962 | 53.5% | | `Nghị quyết` | 26,086 | 16.4% | | `Thông tư` | 16,212 | 10.2% | | `Bản dịch văn bản` | 10,576 | 6.7% | | `Chỉ thị` | 7,908 | 5.0% | | `Nghị định` | 4,690 | 3.0% | | `Thông tư liên tịch` | 3,393 | 2.1% | | `Văn bản hợp nhất` | 1,788 | 1.1% | | `Sắc lệnh` | 980 | 0.6% | | `Công văn` | 708 | 0.4% | | `Luật` | 580 | 0.4% | | `Lệnh` | 376 | 0.2% | ## Lĩnh vực pháp luật · `legal_area` `legal_area` is the canonical Vietnamese label pulled from the ``documentFields[]`` block on each detail-API response (~250 distinct labels in the corpus). Roughly two thirds of the documents are tagged `Chưa phân loại` ("uncategorised") on the source portal; we preserve that literal so consumers can see the gap explicitly instead of nulling it. Top areas below; embedding scatters below show the structural overlap between adjacent areas. | Value | Count | Share | |---|---:|---:| | `Chưa phân loại` | 112,784 | 71.0% | | `Đất đai` | 2,861 | 1.8% | | `Quản lý thuế, phí, lệ phí và thu khác của ngân sách nhà nước` | 2,301 | 1.4% | | `Quản lý thuế, phí và lệ phí` | 1,698 | 1.1% | | `Lĩnh vực giá` | 1,673 | 1.1% | | `Tổ chức- Biên chế` | 1,424 | 0.9% | | `Lao động, tiền lương, tiền công` | 1,373 | 0.9% | | `Quản lý ngân sách` | 1,211 | 0.8% | | `Quản lý ngân sách nhà nước` | 1,052 | 0.7% | | `Ngân sách nhà nước` | 951 | 0.6% | | `Đường bộ` | 942 | 0.6% | | `Chính quyền địa phương` | 869 | 0.5% | | `Công chức, viên chức` | 715 | 0.5% | | `Tài chính khác` | 657 | 0.4% | | `Lâm nghiệp` | 637 | 0.4% | | `Môi trường` | 557 | 0.4% | | `Nông nghiệp` | 513 | 0.3% | | `Thủy lợi` | 473 | 0.3% | | `Chính sách thuế` | 465 | 0.3% | | `Tổ chức cán bộ` | 407 | 0.3% | ## Cơ quan ban hành · Issuing agency Top issuing agencies (top 15). Quốc hội + Chính phủ + Bộ Tài chính + Bộ Tư pháp + ... thường chiếm phần lớn `trung_uong`; các tỉnh chia khá đều phần `dia_phuong`. | Value | Count | Share | |---|---:|---:| | `Bộ Tài chính` | 11,793 | 7.4% | | `Thủ tướng Chính phủ` | 7,849 | 4.9% | | `Chính phủ` | 7,168 | 4.5% | | `UBND Thành phố Hồ Chí Minh` | 5,312 | 3.3% | | `Bộ Nông nghiệp và Phát triển nông thôn` | 2,526 | 1.6% | | `UBND Tỉnh Lào Cai` | 2,497 | 1.6% | | `UBND Tỉnh Nghệ An` | 2,477 | 1.6% | | `UBND tỉnh Bình Phước` | 2,410 | 1.5% | | `Ngân hàng Nhà nước Việt Nam` | 2,198 | 1.4% | | `UBND Thành phố Hà Nội` | 2,123 | 1.3% | | `UBND thành phố Đà Nẵng` | 2,018 | 1.3% | | `UBND tỉnh Bà Rịa - Vũng Tàu` | 1,956 | 1.2% | | `Bộ Giao thông vận tải` | 1,925 | 1.2% | | `UBND Thành phố Hải Phòng` | 1,885 | 1.2% | | `UBND Tỉnh Gia Lai` | 1,765 | 1.1% | ## Năm ban hành · Year of issue Phân bố năm theo `ngay_ban_hanh` (ISO `YYYY-MM-DD`); rỗng nếu cổng không cung cấp được trường này. — Year distribution from `ngay_ban_hanh`; null when the source portal didn't expose the issue date. | Year | Count | |---:|---:| | 2026 | 761 | | 2025 | 9,427 | | 2024 | 7,307 | | 2023 | 5,367 | | 2022 | 5,378 | | 2021 | 5,306 | | 2020 | 4,682 | | 2019 | 4,983 | | 2018 | 4,979 | | 2017 | 5,591 | | 2016 | 6,307 | | 2015 | 5,346 | | 2014 | 5,880 | | 2013 | 5,204 | | 2012 | 4,893 | | 2011 | 4,834 | | 2010 | 4,431 | | 2009 | 4,930 | | 2008 | 5,453 | | 2007 | 6,713 | | 2006 | 6,684 | | 2005 | 5,486 | | 2004 | 5,120 | | 2003 | 4,281 | | 2002 | 3,627 | | 2001 | 3,395 | | 2000 | 2,785 | | 1999 | 3,079 | | 1998 | 2,746 | | 1997 | 2,346 | ## Nguồn nội dung · Body provenance Một văn bản trên `vbpl.vn` có thể có **HTML thân bài** (do API SPA trả về sau khi qua reCAPTCHA) và/hoặc một **tệp đính kèm** (`.pdf` / `.doc` / `.docx`). Pipeline ưu tiên parse tệp khi có, quay về HTML khi không. — A vbpl document may have an inline **body HTML** (returned by the SPA's API after reCAPTCHA) and/or a downloadable **attachment** (`.pdf` / `.doc` / `.docx`). The pipeline prefers parsing the file when present and falls back to the HTML otherwise. | Value | Count | Share | |---|---:|---:| | `body_html` | 147,317 | 92.8% | | `shell_html` | 11,505 | 7.2% | ## Lược đồ bảng `documents` · `documents` schema The parquet has three families of columns: ### Identification + meta | Field | Type | Description | |---|---|---| | `doc_name` / `item_id` | string | Stable document id (= the `--` suffix of the source URL). Mostly numeric (`186739`); legacy docs use `vbpqta_` (Văn bản pháp quy toàn văn) or `vbpqdinhchinh_` (corrigendum). | | `scope` | string | `trung_uong` (central) \| `dia_phuong` (provincial). | | `source` | string | Source host, always `vbpl.vn`. | | `source_url` / `api_url` | string | Deep link back to the portal page / the underlying gateway API. | | `title` | string (nullable) | Document subject after the full title scrub: (1) baseline cleanup (NFC, HTML-entity decode, `"số số"` doubling fix, trailing sentence punctuation), (2) decorative quote stripping at the title boundary and around inline phrases — leading / trailing / paired runs of `"`, `'`, `"`, `'`, `'`, `'` are removed but **word-internal single quotes are preserved** so legitimate Vietnamese / loan-word names survive intact (`Đắk R'lấp`, `M'nông`, `D'Ran`, `Côte d'Ivoire`, `Ea T'ling`, `Đạ M'ri`, …), (3) leading `" số "` boilerplate head peeled (e.g. `"Quyết định số 143/QĐ-KHTC Ban hành Quy chế"` → `"Ban hành Quy chế"`), and (4) every `" [số] [ngày ]"` cross-reference of OTHER documents stripped from anywhere in the body (`"Bãi bỏ Nghị quyết số 84/2018/NQ-HĐND ngày 07/12/2018 của HĐND tỉnh..."` → `"Bãi bỏ của HĐND tỉnh..."`). The Vietnamese noun `Lỗi` / `lỗi` / `loi` ("fault / error") is preserved wherever it appears — it's a legitimate subject word (e.g. `"Lỗi Ban hành Quy chế hoạt động của hội đồng tư vấn mua sắm tài sản"`), not a CMS marker. `null` for the pathological tail where the entire source title was just a bare doc-num token (e.g. `"1938/QĐ-UBND"`); other bibliographic columns still carry the citation handle. | | `doc_type` | string | Self-describing **ASCII snake_case slug** of the Vietnamese doc-type name (`quyet_dinh`, `nghi_dinh`, `thong_tu_lien_tich`, `chi_thi`, `van_ban_hop_nhat`, ...). Auto-derived from `legal_type` via `slugify_vi`; the compact short code (`QĐ`, `NĐ`, `TTLT`, ...) still appears in `so_hieu` itself (`43/2026/NĐ-CP`) and is recoverable via `SLUG_TO_CANONICAL_CODE`. | | `legal_type` | string | Canonical Vietnamese **full name** for the document type (`Luật`, `Nghị định`, `Thông tư`, `Quyết định`, `Nghị quyết`, `Chỉ thị`, …). Round-trips with `doc_type` via `slugify_vi`. | | `legal_area` | string | Legal area / subject domain (`Đất đai`, `Đường bộ`, `Lĩnh vực giá`, …). Defaults to `Chưa phân loại` when the source portal hasn't tagged the doc. | | `so_hieu` | list<string> (nullable) | Document number(s) as a list of canonical short forms (e.g. `["43/2026/NĐ-CP"]`). A small minority of vbpl rows pack several identifiers into one source cell separated by Vietnamese ` và ` ("and") or `,`; those ship as multi-element lists (e.g. `["142/2009/QĐ-TTg", "49/2012/QĐ-TTg"]`). Source-side cruft is stripped: leading legal-type words (`"Nghị quyết số: 528/2018/UBTVQH14"` → `["528/2018/UBTVQH14"]`), trailing annotations (`"109/2005/QĐ-BCA (A11)"` → `["109/2005/QĐ-BCA"]`, `"49/2007/TTLT-BTC-BGD ngày 18/5/2007"` → `["49/2007/TTLT-BTC-BGD"]`), and the legitimate `"Không số"` ("no number") sentinel is preserved. `null` for rows with no usable number. | | `ngay_ban_hanh` | string | Issue date, ISO `YYYY-MM-DD`. | | `year` | int32 | Year extracted from `ngay_ban_hanh`. | | `co_quan_ban_hanh` | string | Issuing agency (e.g. `"Chính phủ"`, `"Bộ Tài chính"`, `"Hội đồng nhân dân tỉnh A"`). | | `trich_yeu` | string | Abstract / summary. | ### Body + stats | Field | Type | Description | |---|---|---| | `markdown` | string (nullable) | NFC-normalised, modern-orthography Vietnamese markdown (page-segmented with `## Page N` headings when parsed from a PDF). Gateway / Word / Next.js scaffolding is stripped: the `Document Content\\n\\nbody { font-family: ... }\\np { margin: ... }` API preamble (~50 % of bodies), Word `` stylesheet dumps, standalone CSS rule blocks gated by property / selector / structural CSS tells, Ant Design `:where(.css-...){ ... }@keyframes ...{ ... }` chains, orphan selector fragments, and malformed inline `` tags. ~90 % of rows lose 1-200 KB of boilerplate; total corpus markdown shrunk by ~1.77 GB / 42 %. **Null** when `body_source == "shell_html"` after the May-2026 live-API recovery (the source genuinely has no body for those legacy IDs); bibliographic metadata (title, agency, so_hieu, ...) is still populated on NULL-markdown rows so consumers retain the citation handle. | | `num_pages` | int32 | Page count from the parser (PDF/DOCX only). | | `num_sections` / `num_paragraphs` / `num_sentences` | int32 | Counts from the structure layer. | | `char_len` | int32 | Character length of `markdown` **after** the junk-strip pass (recomputed at export time so consumers can dedupe on the actual shipped body). | | `text_hash` | string | SHA-256 first-32 hex of `markdown` **after** the junk-strip pass (re-run-stable id; differs from the upstream `extract.jsonl` hash for any row that had scaffolding removed). | | `parser_model` | string | Backend that produced the markdown (`local/pypdf`, `local/markdownify`, `nvidia/nemoretriever-parse`, …). | | `parser_runtime` | string | The configured `parser.runtime` (`local` / `nim` / `hybrid`). | | `body_source` | string | Which source produced the body: `file` (downloaded PDF/.doc/.docx), `body_html` (API-captured), `shell_html` (Next.js shell fallback -- the gateway never delivered a real body; published with `markdown=null` after the May-2026 live-API recovery sweep confirmed the source publishes only metadata for those legacy IDs). | | `parsed_at` | string | ISO 8601 parser timestamp. | ### Hierarchy + entities | Field | Type | Description | |---|---|---| | `structure_json` | string | Full :class:`DocumentStructure` (meta + stats + sections + paragraphs + sentences) as JSON. Includes char-span back-pointers so any unit can be located in `markdown` precisely. | | `extracted_json` | string | Generic NER + statute-link extraction (entities, relations, statute_refs) as JSON. | | `file_paths_json` | string | Downloaded attachments as JSON list of `{file_url, file_name, file_type, local_path}`. | Quick load: ```python import json from datasets import load_dataset ds = load_dataset("tmquan/vbpl-vn", split="train") row = ds[0] print(row["doc_type"], row["so_hieu"], row["ngay_ban_hanh"]) # e.g. "quyet_dinh 143/QĐ-KHTC 2018-01-29" print(row["title"]) # e.g. "Ban hành Quy chế quản lý ngân sách ngành Tư pháp" # (the redundant "Quyết định số 143/QĐ-KHTC " head has been # stripped; recombine via f"{row['legal_type']} số {row['so_hieu']} {row['title']}" # if your downstream pipeline still wants the full original.) structure = json.loads(row["structure_json"]) for sec in structure.get("sections", []): print(sec["kind"], sec["label"]) ``` ## Companion stages · `embed` + `reduce` Alongside the default `documents-*.parquet` shards (one row per document, with text + structure), the repo also carries the **embed** and **reduce** pipeline outputs as separate parquet bundles. Both join back to the `documents` table on the `doc_name` primary key. Only the **embeddable** rows (147,317 after dropping NULL-markdown docs) appear in these stages. ### `embed-*.parquet` — dense vectors 15 shards (~93 MB each, ~1.33 GB total, 10 000 rows per shard, deterministic `doc_name` ordering). Schema mirrors the `anle.toaan.gov.vn` corpus's embed stage exactly so cross-corpus joins are straightforward: | Field | Type | Description | |---|---|---| | `doc_name` | string | Join key back to `documents-*.parquet`. | | `text_hash` | string | SHA-256 of the post-normalisation `markdown` (stable across re-runs). | | `embedding` | list<float64> | **2048-D** dense vector from `nvidia/llama-nemotron-embed-1b-v2` (default; other models give other dims). | | `embedding_dim` | int64 | Length of `embedding` (denormalised for fast filtering, always `2048` in this release). | | `embedding_model_id` | string | Model slug as the embedder backend reports it. | | `embedding_text_hash` | string | SHA-256 of the exact text fed to the embedder (differs from `text_hash` when sliding-window chunking applies). | | `embedding_chunks_used` | int64 | Number of windows mean-pooled into the final vector (1 when the doc fits in one window). | | `embedding_chunking` | string | Chunking strategy: `off` / `sliding` / `sentence`. | ### `reduce-*.parquet` — 2-D projections + cluster ids 15 shards (~0.5 MB each, ~7 MB total). PCA + t-SNE + UMAP run with `cfg.reducer.n_components=2` so the `*_z` columns that existed in the on-disk per-doc shards are dropped here. HDBSCAN cluster ids land in `[-1, N]` (`-1` is the noise bucket). | Field | Type | Description | |---|---|---| | `doc_name` / `text_hash` | string | Join keys back to `documents-*.parquet` and `embed-*.parquet`. | | `pca_x` / `pca_y` | float64 | 2-D PCA projection of the 2048-D embedding. | | `tsne_x` / `tsne_y` | float64 | 2-D t-SNE projection. | | `umap_x` / `umap_y` | float64 | 2-D UMAP projection (the one used in the scatter PNGs above). | | `cluster_id` | int64 | HDBSCAN cluster label; `-1` is the noise / unclustered bucket. | Quick load (each stage is a `data_files` glob; the default `load_dataset("tmquan/vbpl-vn")` still resolves to the `documents` config): ```python from datasets import load_dataset embed = load_dataset( "tmquan/vbpl-vn", data_files="embed-*.parquet", split="train", ) print(embed[0]["doc_name"], len(embed[0]["embedding"])) # e.g. "100000 2048" reduce = load_dataset( "tmquan/vbpl-vn", data_files="reduce-*.parquet", split="train", ) print(reduce[0]["doc_name"], reduce[0]["umap_x"], reduce[0]["cluster_id"]) # e.g. "100000 1.7142 -1" ``` To join embed-stage vectors back to the document metadata, do the join client-side on `doc_name`: ```python import pandas as pd docs = load_dataset("tmquan/vbpl-vn", split="train").to_pandas() embed = load_dataset("tmquan/vbpl-vn", data_files="embed-*.parquet", split="train").to_pandas() joined = docs.merge(embed, on="doc_name", how="inner") # joined now has the title / so_hieu / markdown / ... columns # next to the 2048-D embedding vector for every embeddable doc. ``` ## Trực quan hoá embedding · Embedding visualization Mỗi điểm là một văn bản pháp luật; toạ độ là vector embedding 2048-D từ `nvidia/llama-nemotron-embed-1b-v2` chiếu xuống 2D bằng UMAP, cụm bằng HDBSCAN. Sáu mặt phân hoạch: `scope`, `doc_type` (mã ngắn), `legal_type` (tên đầy đủ), `legal_area` (lĩnh vực pháp luật), `year`, `cluster_id`. Các nhãn tail-end (sau top-18) được dồn vào nhóm *Khác / Other* màu xám để chú giải đọc được. — Each dot is one legal document; coordinates are the 2D UMAP projection of a 2048-D embedding from `nvidia/llama-nemotron-embed-1b-v2`, with HDBSCAN cluster ids. Six facets: `scope`, `doc_type` (canonical short code), `legal_type` (canonical full name), `legal_area` (subject domain), `year`, `cluster_id`. Tail-end labels beyond the top 18 are collapsed into a grey *Khác / Other* bucket to keep the legend legible. The published reducer parquet still carries `tsne_x` / `tsne_y` columns next to `umap_*` for consumers who want to render t-SNE themselves. ### UMAP colored by `scope` ![UMAP colored by `scope`](./embedding-scope-umap.png) ### UMAP colored by `doc_type` ![UMAP colored by `doc_type`](./embedding-doc-type-umap.png) ### UMAP colored by `legal_type` ![UMAP colored by `legal_type`](./embedding-legal-type-umap.png) ### UMAP colored by `legal_area` ![UMAP colored by `legal_area`](./embedding-legal-area-umap.png) ### UMAP colored by `year` ![UMAP colored by `year`](./embedding-year-umap.png) ## Cách thu thập + chuẩn hoá · How the corpus was built The crawler is a six-stage pipeline (`harvest` → `detail` → `parse` → `extract` → `embed` → `reduce`) that walks vbpl.vn's public sitemap chain (32 shards, ~160 K URLs total), drives a headless Chromium tab against each detail page so Google's invisible reCAPTCHA v2 mints the per-session Bearer token (the SPA's `/api/qtdc/public/doc/...` gateway is otherwise inaccessible), intercepts the resulting authenticated XHRs, downloads any `.pdf` / `.doc` / `.docx` attachment, and routes the body through: 1. **Parse** -- pypdf for PDFs, docx2txt for `.docx`, an `antiword` / `catdoc` / `libreoffice` subprocess fallback for legacy `.doc`, `markdownify` for HTML bodies returned inline. 2. **Vietnamese normalisation** -- ftfy NFC + tone-mark canonicalisation (`Toà → Tòa`, `hoà → hòa`, `thuỷ → thủy`) + PDF whitespace cleanup. Every regex / segmenter downstream then sees a single canonical orthography. 3. **Generic + structure extractor** -- regex / dictionary NER + Vietnamese statute linker (`Điều N khoản M Luật ...`, dates `dd/MM/yyyy`, courts, agencies, document numbers) + a hierarchical `DocumentStructure` (sections / paragraphs / sentences with back-pointers). 4. **Embed** -- `nvidia/llama-nemotron-embed-1b-v2` (2048-D) over the normalised markdown; sliding-window mean pool when a doc exceeds the model's native context window. 5. **Reduce** -- PCA + t-SNE + UMAP on the embedding matrix + HDBSCAN cluster ids. cuML on a GPU worker; sklearn / umap-learn / hdbscan otherwise. All five layers are deterministic and re-runnable; re-running any stage with the same `--limit` is a no-op (each stage skips already-produced outputs). ### Body-recovery rerun (May 2026) A targeted retry against the public gateway `https://vbpl-bientap-gateway.moj.gov.vn/api/qtdc/public/doc/` recovered **3,849 / 15,351** previously bodyless documents (the gateway now publishes real `documentContent.content` HTML for that slice even though our cached SPA fetch came back empty in 2026-04). The remaining ~11.5K are genuinely bodyless on the official vbpl source (legacy / cancelled documents the publisher no longer carries a full text for) and ship in this dataset with `markdown=null`. Bibliographic metadata (title, agency, document number, issue date, ...) is still populated on those rows so consumers retain the citation handle even when no body is available. Per-doc embedding parquets for the NULL-markdown rows are dropped from the embedding corpus entirely so the reducer fits on the embeddable rows only (147,317-row corpus). Captured: `2026-05-21T17:50:24.822753+00:00`. ## Nguồn · Source * Portal: * Backend gateway: `vbpl-bientap-gateway.moj.gov.vn` * Publisher: Ministry of Justice of Vietnam (Bộ Tư pháp) * Sitemap: ## Giấy phép · License Văn bản gốc được Bộ Tư pháp công bố trên cổng thông tin công cộng (`Allow: /` trong `robots.txt`). 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 Ministry of Justice on a public portal (its `robots.txt` allows `/` and disallows only `/api/`). 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** (Cơ sở dữ liệu Quốc gia về pháp luật, Bộ Tư pháp Việt Nam): ```bibtex @misc{vbpl_2026, title = {Vietnamese National Legal Database (vbpl.vn)}, author = {TMQuan}, year = {2026}, howpublished = {\url{https://huggingface.co/datasets/tmquan/vbpl-vn}}, note = {Document-level mirror with a hierarchical structure layer (DocumentMeta + Section + Paragraph + Sentence) over Vietnam's National Legal Database, central + provincial scope.} } @misc{vbpl_moj_2026, title = {Vietnamese National Legal Database}, author = {{Cơ sở dữ liệu Quốc gia về pháp luật}}, year = {2026}, howpublished = {\url{https://vbpl.vn/}}, note = {Official portal for Vietnam's National Legal Database (laws, ordinances, decrees, circulars, decisions, ...) at central and provincial levels, published by the Ministry of Justice (Bộ Tư pháp).} } ```