V4ldeLund/faroese-logir-fo

Logir collection — 28 September 2026

Collected the public HTML legal texts from Logir.fo. The official-document exclusion in §9 of the Faroese Copyright Act is the recorded rights basis; this is not a CC0 dedication or a blanket licence for the website.

Results

Output	Documents	Characters	Llama-3 tokens
faroese.parquet	5,172	38,846,754	16,588,291
danish.parquet	2,148	42,623,383	14,638,673
Combined corpus	7,320	81,470,137	31,226,964

Tokenizer: AI-Sweden-Models/Llama-3-8B-instruct, with the existing Dynaword counting helper and its default special-token handling. Matching gzip-compressed JSONL files retain nested paragraph and table structure.

Coverage and exclusions

• Discovered all 7,420 records in the live search index: 5,271 Faroese and 2,149 Danish.
• Successfully parsed HTML responses: 7,420.
• Failed records: 0; unprocessed URLs: 0.
• Pages whose full text is a separate file, excluded from the main corpus: 100 (inventory).
• Exact text duplicates removed: 0 (mapping).
• Documents isolated for explicit third-party copyright indicators: 0 (review file).
• Available text from incomplete pages is preserved in the inventory, including one circular introduction; it is not counted as a complete legal document.
• Navigation, duplicated chapter menus, images and embedded objects were removed. Statutory text and textual annexes were retained.
• The robots-disallowed /GetFile and /Content routes were not downloaded. This limits PDF-only laws and earlier consolidated snapshots served by the historical-version viewer.
• 285 records advertise historical-version indexes; those index URLs are preserved in metadata.
• The public search grid supplied discovery after the advertised crawler index returned HTTP 500. Pagination overlap was deduplicated and the language totals reconciled with the independent all-language count.

Verification

Validation: passed. Compared all 7,320 exported legal bodies with the saved source HTML after the documented exclusions and text normalization, ignoring layout whitespace. Also checked hashes, lengths, duplicate removal, language splits, source scope, coverage accounting, and Parquet/JSONL agreement.

Sixteen offline regression tests cover extraction boundaries, inline words, tables and captions, metadata-only pages, copyright-notice quarantine, URL scope, and pagination.

Legal status of retained texts

Status in Logir	Documents
Áður galdandi	3,045
Galdandi	3,872
Søgulig	398
Í koming	5

Document types

Type	Documents
Anordning	276
Anordningsbekendtgørelse	11
Bekendtgørelse	1,293
Lovbekendtgørelse	55
Forordning	4
Cirkulære	2
Fráboðan	102
Kirkjulig fyriskipan	1
Kunngerð	3,034
Kundgørelse	2
Lov	305
Løgtingslóg	1,905
Lagtingslov	163
Leiðbeining	46
Løgtingslógarkunngerð	3
Midlertidig bestemmelse	12
Norske Lov	20
Plakat	4
Politivedtægt	2
Rundskriv	79
Tingskipan	1

Reuse and provenance

Each row records the exact source URL, retrieval time, document type, issuing authority, date, language, validity status, source-response hash and extracted-text hash. Language is the source search classification and does not guarantee that every quoted passage is in that language. Historical texts and amendments remain in the corpus with status labels.

Independent contributions remain subject to §9(2). Explicit notices are screened automatically; absence of a notice does not establish the status of every embedded contribution. The corpus is restricted to recognized official legal-document bodies, and raw website response caches are evidence files, not part of the extracted-text release.

Parser and rerun instructions · Machine-readable report · Validation details · Tokenizer counts

Sources: Faroese Copyright Act, §§9 and 49, Logir database description, robots.txt. Logir states that the gazetted text is legally authoritative.

Load the dataset

from datasets import load_dataset

faroese = load_dataset("V4ldeLund/faroese-logir-fo", "faroese", split="train")
danish = load_dataset("V4ldeLund/faroese-logir-fo", "danish", split="train")
combined = load_dataset("V4ldeLund/faroese-logir-fo", "all", split="train")

Faroese is the default configuration. The train label is a distribution split; this collection does not define evaluation splits.

Parquet stores nested fields such as blocks, references, and metadata as JSON strings. Use json.loads to decode those fields, or download the structured Faroese JSONL, Danish JSONL, or combined JSONL.

The parser and its reused crawler helpers are included under scrapers/. The published files are listed with SHA-256 hashes in release_manifest.json. Raw HTML caches, request logs, and the local SQLite state are not part of this release.