Statskontoretdatalabb/Regeringsuppdrag

Swedish government assignments from Athena

Overview

This dataset brings together Swedish government assignments (regeringsuppdrag) with publication metadata, receiving organisations, extracted text, Markdown, structured Docling JSON and original PDFs. It supports document exploration, text analysis and comparison with the original documents.

Each row represents one content file, not one publication. A publication with several files produces several rows with the same publication_id and different id values. Publication metadata is repeated on those rows.

The intended full export includes all available government assignments imported into the selected Athena instance. Availability in Athena does not establish that every assignment ever published by the government is included.

Sources and coverage

Original publications and PDFs come from Regeringen.se. Athena's importer uses the g0v.se government assignment index to discover publications and attachments. It imports metadata and PDFs, matches receiving organisations and uses Docling to extract content.

Use context_url for the publication page and url for the source PDF. The export reuses Athena's stored originals, bundled under pdfs/. Coverage depends on source availability, import results and export selection. The current command takes explicit publication IDs; exporting the full collection requires selecting all intended publications. Missing extracted content remains null. The published column describes source publication dates, not import dates.

Test upload — not the full dataset

This pull request contains a development sample exported on 2026-11-10: 3 publications and 3 PDF files, with the current 19-column schema. The sample selects publications with available Docling JSON to exercise the publication workflow. It is not representative of the full collection. Parquet size: 42,938 bytes. Total bundled PDF size: 413,734 bytes. Local validation completed successfully with no errors. The publication count matches the selected publication IDs. A full export will be prepared separately.

Files and default loading

README.md
regeringsuppdrag.parquet
pdfs/<content-id>.pdf

Rows are written in batches to one Parquet file using ZSTD level 3 compression. Rows are sorted by publication date, publication ID, file sequence and content ID. Missing dates and sequences sort last. PDFs remain separate so readers can use extracted content without downloading originals. After export, the command saves validation.json with a machine-readable validation summary, including failures. Validation errors produce a nonzero command exit status. Validation reports stay local and are not uploaded with the dataset. The manually maintained README is preserved during export and validation.

The configuration above selects only regeringsuppdrag.parquet, excluding PDFs and JSON reports from dataset loading. The split name train is a loading convention, not a claim that this is a curated machine-learning training set. This follows the Hugging Face manual configuration format.

Fields

The export contains the 19 columns listed below.

id, publication_id and author are non-nullable in the Arrow schema. Other columns are nullable, although validation requires PDF references and a recipient list. An empty recipient list is valid. Missing extracted content does not imply that the original document is empty.

“JSON string” means serialized JSON in a Parquet string column: use json.loads() to access its structure. Recipient and registry-number lists are native Parquet lists. published is a calendar date without a time or timezone.

Column	Type	Meaning
id	string, required	Content UUID, unique row ID and PDF basename.
publication_id	string, required	Parent publication UUID.
title	string	Publication title.
author	string, required	Publishing organisation, always Regeringskansliet.
published	date32	Source publication date without time or timezone.
registry_numbers	list of strings	Identifiers using the registry:regeringskansliet scheme.
recipients	list of strings	Sorted unique organisation names with the subject role.
context_url	string	Public publication page.
content_title	string	Individual file title.
sequence	int32	File position within its publication.
is_primary	boolean	Whether the file is marked as primary content.
file_name	string	Original filename or attachment label.
url	string	Public source PDF URL.
pdf_path	string	Relative path to the bundled PDF.
pdf_md5	string	MD5 of the PDF bytes, reused from Athena's file_hash; null when unavailable.
text	string	Extracted plain text.
markdown	string	Extracted Markdown.
json	JSON string	Structured extraction output, normally a DoclingDocument.
docling_version	string	Docling extraction package version, when known.

The public schema intentionally omits internal import and processing metadata, raw source snapshots, duplicated IDs and full database relationships. pdf_md5 preserves the existing source-file hash for comparing PDF bytes across exports without downloading previous PDFs. It does not cover text extraction or publication metadata and is not a security guarantee against deliberate tampering. docling_version is retained to identify the extraction package version per file. All files are included; users can filter is_primary to select main content. Several files within a publication may be primary, and the flag is not a guarantee that a file is the original or latest decision.

Registry numbers do not retain the source's “delvis” (is_partial) qualifier. A compact representation of that qualifier and a separate PDF checksum manifest are deferred options for a future release (field review: #370). Neither is included in this export. Athena retains its internal metadata and hash checks.

Reading the data

In an Athena checkout, start Python with uv run --group export python. Run the examples from the export directory, or change root to that directory.

Markdown without PDFs

from pathlib import Path
import pyarrow.parquet as pq

root = Path(".")
markdown_table = pq.read_table(
    root / "regeringsuppdrag.parquet",
    columns=["id", "publication_id", "title", "markdown"],
)
print(markdown_table.slice(0, 1).to_pylist())

Only the Parquet file is needed. For a remote release, download that file alone; downloading the entire repository would also download PDFs. Column selection does not reduce the size of a Parquet file already downloaded in full.

Optional original PDF

from pathlib import Path
import pyarrow.parquet as pq

root = Path(".")
row = pq.read_table(
    root / "regeringsuppdrag.parquet",
    columns=["id", "url", "pdf_path"],
).slice(0, 1).to_pylist()[0]
pdf = root / row["pdf_path"]
print("Open this PDF in a reader:", pdf.resolve())
print("Source URL:", row["url"])

This second example requires downloading the referenced PDF too.

Validation and limitations

Validation checks schema, row-ID uniqueness, required IDs, recipient types, Docling JSON structure, PDF paths, file existence and PDF headers. The exporter checks stored PDF bytes against Athena’s internal MD5 when available. Standalone validation does not currently compare pdf_md5 with the bundled PDF and cannot detect arbitrary later changes to a PDF that retains its header. Missing extracted content is reported separately; an empty export or invalid PDF reference fails validation. Counts distinguish rows, publications, PDF references and unique PDF files. Source completeness can only be checked when expected source counts are supplied.

Earlier development used a limited sample to save processing time. This schema covers all regeringsuppdrag, not a fixed-size sample. Generate and validate a fresh export before publishing, and assess extraction quality against original PDFs.

OCR may misread characters. Tables, reading order, headers and footnotes may be reconstructed incorrectly. Recipient matching can miss or misidentify organisations. A valid DoclingDocument structure does not prove that extracted text is accurate; compare with original PDFs when accuracy matters.

Database reads use separate requests, not a transactional snapshot. Produce releases from a stable instance. Athena UUIDs may change after a database reset and reimport.

Reuse terms and publication scope

No dataset-wide reuse licence has been established for this collection. This card does not grant permission to redistribute source documents or third-party material within them. Verify source terms and attribution requirements before public release; no licence is claimed in the card metadata.

The public schema contains the 19 fields listed above. Internal source metadata, importer and processor implementation details are excluded. pdf_md5 reuses Athena's stored PDF hash; docling_version identifies the extraction version. Preserve source links when preparing a release.

Maintenance

This card is maintained manually, following Athena issue #279. It describes the dataset design. The sample results above come from the test export on 2026-11-10, not from a full release.

When exporting the full collection, refresh coverage and validation information using that export's results, check this field table against its schema, and copy the reviewed card into the export directory as README.md. Keep sample and release statistics distinct. Export and validation must not overwrite manual card edits.