MeerDevelopment/dutch-tax-text-corpus

Dutch Tax Text Corpus (2026-10-19)

Public Dutch tax and customs texts, cleaned and split into passages, with the source URL, crawl time, tax year and reuse terms on every record. The corpus combines website text from the Belastingdienst, its Kennisgroepen, Rijksoverheid, Government.nl and the Douane with 31 tax laws and 1,605 Staatscourant publications.

It is meant as a ready base for Dutch tax search and RAG systems, retrieval benchmarks and the evaluation of language models on Dutch tax questions. It was compiled by MeerDevelopment from a crawl on 2026-10-19.

This is an unofficial snapshot. It is not endorsed by or affiliated with the Belastingdienst, the Douane or any other source authority, and it is not tax or legal advice. Tax rules change every year, so always check the original source.

Contents

Configuration	File	Rows	Contents
main_chunks (default)	main/chunks.parquet	37,345	Text passages from websites, laws and Staatscourant
main_documents	main/documents.parquet	9,842	Metadata per document, with its chunk IDs
douane_chunks	douane/chunks.parquet	2,249	Text passages from douane.nl
douane_documents	douane/documents.parquet	1,391	Metadata per Douane document

The documents tables hold metadata only. The text lives in the chunks tables, and a document's chunks are the passages that survived deduplication, so they do not always add up to the complete original page. The Douane texts are a separate configuration because their reuse terms differ (see below). Each configuration has a single train split; it is not a curated train/test division.

Source	Documents	Chunks	Coverage	Reuse basis
belastingdienst.nl (NL and EN)	4,228	6,195	Website, crawled 2026-10-19	CC0 1.0
kennisgroepen.belastingdienst.nl	2,694	6,182	Standpunten and news	CC0 1.0
rijksoverheid.nl	946	1,492	Tax topics	CC0 1.0, except texts marked as copyrighted
government.nl	147	206	Tax topics in English	CC0 1.0, except texts marked as copyrighted
Legislation (BWB)	222	4,731	31 laws, versions in force 2020 to 2026	Public act, no copyright (Auteurswet art. 11)
Staatscourant	1,605	18,539	Tax decisions published 2008 to 2026	Public act, no copyright (Auteurswet art. 11)
douane.nl	1,391	2,249	Website	Reuse allowed with source naming, see below

Usage

from datasets import load_dataset

repo = "MeerDevelopment/dutch-tax-text-corpus"
chunks = load_dataset(repo, "main_chunks", split="train")
documents = load_dataset(repo, "main_documents", split="train")

cc0_only = chunks.filter(lambda row: row["license"] == "CC0-1.0")

doc_by_id = {doc["doc_id"]: doc for doc in documents}
chunk = chunks[0]
print(chunk["text"][:200])
print(doc_by_id[chunk["doc_id"]]["url"], chunk["crawl_timestamp"])

Use douane_chunks and douane_documents for the customs texts. Filter on license to select one reuse basis: CC0-1.0 for the website texts, LicenseRef-NL-Public-Acts for laws and Staatscourant.

Licence and attribution

There is no single licence for the whole dataset; each record carries its own license, license_url, license_basis, attribution and integrity_terms. In short:

• Website text from belastingdienst.nl, kennisgroepen, rijksoverheid.nl and government.nl (14,075 chunks) is published by those sites under CC0 1.0. Images and other components with their own notice are excluded from that and are not in this dataset.
• Laws and Staatscourant publications (23,270 chunks) have no copyright under article 11 of the Dutch Copyright Act. That is a statutory basis, not a CC0 grant.
• Douane text (2,249 chunks) may be reused if you name the Douane as the source and do not change its meaning.

Details and links to the original terms are in LICENSE.md. Reusers must not suggest that a source authority has reviewed or endorsed their work.

Fields

Fields in both tables:

Field	Type	Meaning
source	string	Source key, such as belastingdienst or bwb
url, fetched_url	string	Canonical source URL and the URL actually fetched
title, lang, topic	string	Title, language (nl or en) and tax topic
raw_format	string	Format of the source (HTML or XML); raw files are not included
as_of	string	Date from the source: modification date, publication date or start of validity
crawl_timestamp	timestamp (UTC)	When the text was downloaded
crawl_timestamp_provenance	string	Which crawl record supplied the timestamp
tax_year, tax_years	int32, list of int32	Tax year(s) the text applies to, or null
license, license_url, license_basis, attribution, integrity_terms	string	Reuse terms per record

Documents only: doc_id, doc_type, valid_to and bwb_id (string), text_chars and retained_chunk_count (int64), tax_relevant (bool) and chunk_ids (list of string).

Chunks only: chunk_id, doc_id, breadcrumb, text and bwb_id (string), doc_ids (list of string) and n_versions (int64).

Notes on some fields:

• tax_year comes from the law version for BWB. For the other sources it is guessed from the title and URL, so treat it as a hint.
• as_of is not the crawl date; use crawl_timestamp for that.
• tax_relevant is an automatic guess. Pages marked false (1,642 documents) are included anyway, and the field is null for laws and Staatscourant.
• A law passage that is identical across versions is stored once. Its doc_ids lists every version it belongs to and n_versions counts them, so use doc_ids, not just doc_id, to rebuild a version.
• chunk_ids gives membership, not a guaranteed reading order. text_chars is the length of the processed document before deduplication.

How it was made

The websites were crawled on 2026-10-19 and the main text was extracted from each page, without menus, footers, images or embedded media. Laws came as XML from the official BWB repository and Staatscourant publications from the open overheid.nl repository. Text was split into passages and near-duplicate passages were removed; documents with no passages left were dropped, so duplicate pages and their alternative URLs are not listed. Passage text is otherwise unchanged.

Left out on purpose:

• all PDF files, including brochures and explanatory notes;
• court rulings;
• business.gov.nl and ondernemersplein.overheid.nl, whose terms forbid redistribution;
• 39 further pages flagged in a rights and privacy audit: direct phone numbers of staff, estate notices naming private persons, third-party licensed text, password-protected pages, pages from other hosts and mixed-language pages.

The four Parquet files are checked against a SHA256 manifest of the verified export.

Limitations

• The privacy and rights screening was automated, with manual review of samples. Names of officials in their public role remain. The dataset may still contain personal data or third-party material that the screening missed.
• Passages can lack context from the surrounding page, and extraction or deduplication can make them incomplete.
• This is a snapshot. Rates, thresholds and rules change, and older Staatscourant decisions may have been withdrawn or replaced.

Citation

@misc{meerdevelopment2026dutchtax,
  title        = {Dutch Tax Text Corpus (2026-10-19)},
  author       = {{MeerDevelopment}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/MeerDevelopment/dutch-tax-text-corpus}},
  note         = {Snapshot crawled on 2026-10-19}
}

Corrections and takedown

Send the document or chunk ID, the source URL and the reason to zander@meerdevelopment.nl, and the record will be corrected or removed.