Applied the per-shard robots.txt masks (train-*.allowed.zst, from PR #1) to the Parquet shards: kept the 138,385 rows with allowed=true, dropped 32,121 (mainly nrk.no 22,051, yle.fi 7,525, radio.nrk.no 1,378, vaalit.yle.fi 751, digitalarkivet.no 151). Row order, schema and compression unchanged; survivors re-sharded at 16,000 rows/shard. The per-shard mask files are removed because they were positionally aligned to the old 11 shards and would no longer match; robots_stats.tsv is kept.
170,506 documents / 71.9M words / 55.0M Sámi words (76.5%).
PDF extraction flattens tables and dot leaders onto single lines, which distorted word counts, language identification and the corpus n-grams (`a b¹ b² b³ a` came from one answer-key grid repeated 40 times).
- Nothing now repeats more than three times: identical consecutive lines, repeated token-cycles inside a line, and runs of a punctuation character are truncated to three copies. Letters and digits are untouched, so numbers like 1000000 survive. 1,368 documents, 422,504 words.
- Four documents also had a single collapsed-table line removed outright (30,114 words); their measured Sámi share rose sharply once the noise was gone (0.375 -> 0.918, 0.359 -> 0.915, 0.316 -> 0.768).
- Blank forms shrink to their real content (an 8,126-word voter-list form is 66 words of heading plus hundreds of empty rows).
- row order unchanged; the rule is now part of the extraction cleaner.
- columns: sha256_id, text, source, url, languages, sa…
170,506 documents / 72.4M words / 55.1M Sámi words (76.2%).
- Binary (PDF/office) documents now need >= 5% Sámi when over 1,000 words, and >= 2% when shorter: 474 further documents removed (8.5M words, 266k Sámi). Their Norwegian bulk was dominating the corpus n-grams; the file documents are now 66.5% Sámi (was 50.2%) and contribute 3.8M Norwegian words instead of 8.2M.
- Fixed a text defect: a SOFT HYPHEN at a line end had its character removed but the newline kept, leaving the word split in two ('ram\nmene' instead of 'rammene'). 13 occurrences in 6 documents repaired; the extraction code now joins these unconditionally, as a discretionary hyphen always should be.
- Sámi share 68.5% -> 76.2%; row order unchanged; web/sikor/wikipedia rows untouched.
- columns: sha256_id, text, source, url, languages, saami_share
170,980 documents / 80.9M words / 55.4M Sámi words (68.5%).
Removed 368 binary documents over 1,000 words whose Sámi share is under 2% (6.0M words, but only 88,921 Sámi words — 0.16% of the corpus's Sámi). These are dominated by their Norwegian/Finnish/English body: a 123k-word Springer volume, Finnish doctoral theses, the English edition of Sámi logut muitalit. Sámi share rises 63.9% -> 68.5%.
- row order unchanged; web/sikor/wikipedia rows untouched
- columns: sha256_id, text, source, url, languages, saami_share
171,348 documents / 86.9M words / 55.5M Sámi words (63.9%).
The earlier binary extraction discarded every non-Sámi page, so documents lost their context and their true Sámi share was unknowable. The new extraction keeps all pages:
- 6,742 PDF/office documents replaced with their full text (19M -> 39.2M words)
- 1,107 binaries that are under 1% Sámi removed outright (20.6M words), e.g. a 297k-word report, a 223k-word NOU, a 200k-word Finnish Bible
- Sámi word count rises slightly (54.6M -> 55.5M) while the share falls (80.3% -> 63.9%), because kept documents now carry their non-Sámi context
- row order unchanged; web/sikor/wikipedia rows untouched
- columns: sha256_id, text, source, url, languages, saami_share
172,455 documents / 68.5M words, shuffled (seed 42).
- Recovered article headlines for regjeringen.no articles missing them (~1,600; h1/og:title, dropping the '- regjeringen.no' <title> suffix).
- Added a frequency guard so a meta-description repeated across many pages (regjeringen.no's site-wide blurb, 521x) is treated as boilerplate and not inserted as an article lead — applies to all recovered hosts.
- (prev rounds) nrk+avvir+yle title recovery, stream dedup, heading demotion, corrected no-Sámi filter.
- columns: sha256_id, text, source, url, languages, saami_share
172,439 documents / 68.6M words, shuffled (seed 42).
- Recovered article headlines for yle.fi articles missing them (~8,000). yle's <h1> is the section name ('Sápmi'), so the headline is taken from og:title, with the '| Tulospalvelu | Aluevaalit' / '| Yle' suffixes and site-name titles stripped; matching is robust to Skolt apostrophe/hyphen orthography so an already-present title is never duplicated.
- (prev rounds) nrk+avvir title recovery, stream dedup, heading demotion, corrected no-Sámi filter.
- columns: sha256_id, text, source, url, languages, saami_share
172,490 documents / 68.3M words (80.2% Sámi), shuffled (seed 42).
- Recovered headline (<h1>) and perex (<meta description>) for 13,005 NRK articles where the LLM extraction had begun at the article body and dropped them (+~302k words); collapsed the 59 URL-variant duplicates this revealed.
- (prev round) NRK stream dedup, prose-heading demotion, corrected no-Sámi filter.
- columns: sha256_id, text, source, url, languages, saami_share
172,457 documents / 68.1M words (80.3% Sámi), shuffled (seed 42).
- NRK recirculation-stream removal reworked (sentence-level, blanket) so unrelated bulletins no longer leak into articles; revealed/collapsed the duplicate crawls the stream had masked.
- 18,811 prose paragraphs wrongly marked as headings demoted to text.
- leftover HTML tags stripped; PDF line-break hyphenation repaired.
- no-Sámi filter corrected: keep any doc with >=1 Sámi sentence or a Sámi-distinctive character; drop only genuinely foreign docs.
- columns: sha256_id, text, source, url, languages, saami_share
Text-only release: the raw HTML/XML/wikitext `original` and its `original_format` are no longer shipped, cutting the download from 1.79 GB to 190 MB. Same 175,366 rows and same shard filenames.
Columns: sha256_id, text, source, url, languages, saami_share
Text re-extracted with an improved prompt and reasoning enabled, plus 8,108 documents extracted from PDFs and office files.
- 175,366 documents / 68.5M words (was 186,000 / 53.0M)
- adds 21.2M words of PDF/office text (original_format: pdf, docx, ...)
- NRK live-blog bulletins repeated across articles removed (keep-first)
- documents with no Sámi at all removed (7,764)
- server-log and data-table dumps removed (13)
- duplicate sha256_id rows fixed (1,403 -> 0)
- languages/saami_share recomputed on the new text
