Goader/ukrainian-news-2026

Ukrainian News 2026

Ukrainian-language news articles from 20 national outlets, published between 1 January and 28 August 2026. Extracted body text plus metadata.

Two configs. deduplicated is the default — near-duplicates removed, which is what you want when mixing this with an already-deduplicated pretraining corpus. raw is the original release, unchanged.

	deduplicated (default)	raw	train-mixin
Documents	419,204	429,427	386,477
Characters	0.97B	1.01B	0.88B
Tokens	~200M	208M (measured with lapa-llm/tokenizer)	~181M
Date range	2026-01-01 → 2026-08-28	same	strictly in-window
Size on disk	0.63 GB, zstd	0.94 GB, zstd	0.59 GB, zstd

All three share the schema, the window and the outlets; language is Ukrainian only throughout.

deduplicated and raw differ on deduplication alone, so anything you could do with one you can do with the other. train-mixin is a purpose-built subset — see below before using it for anything other than the purpose it was cut for.

License

These data are released under this licensing scheme:

• We do not own any of the text from which these data have been extracted. Every article remains the property of the outlet that published it. Treat the article text as all-rights-reserved.
• We license the actual packaging, the metadata and the annotations of these data under the Creative Commons CC0 license ("no rights reserved"). That covers the Parquet layout, the schema, the extraction and language-detection pipeline, and the derived fields we computed: lang, fingerprint, text_len, n_paragraphs, id.

For any use beyond private research, consult each outlet's own terms — they are linked in the table below. We make no representation that any particular use is permitted, and that determination is yours.

Sources

key	outlet	site	documents	tokens	median chars
24tv	24 Канал	24tv.ua	47,432	26.9M	2,334
ukrinform	Укрінформ	ukrinform.ua	45,074	16.4M	1,339
glavcom	Главком	glavcom.ua	36,868	19.1M	2,058
pravda	Українська правда	pravda.com.ua	32,457	20.1M	1,643
obozrevatel	Обозреватель	obozrevatel.com	31,580	18.8M	2,470
rbc	РБК-Україна	rbc.ua	28,447	14.4M	2,181
censor	Цензор.НЕТ	censor.net	26,957	8.7M	1,429
interfax	Інтерфакс-Україна	interfax.com.ua	25,289	8.7M	1,356
focus	Фокус	focus.ua	23,240	14.5M	2,576
nv	НВ	nv.ua	22,586	11.2M	1,856
lb	LB.ua	lb.ua	21,740	8.7M	1,272
ukranews	Українські Національні Новини	ukranews.com	16,283	5.8M	1,502
unian	УНІАН	unian.ua	15,102	8.3M	2,397
zaxid	Zaxid.net	zaxid.net	13,123	6.8M	1,993
liga	Ліга	liga.net	11,414	4.6M	1,835
slovoidilo	Слово і Діло	slovoidilo.ua	9,115	4.0M	1,959
hromadske	Громадське	hromadske.ua	8,824	4.0M	1,702
suspilne	Суспільне	suspilne.media	5,859	3.1M	2,007
epravda	Економічна правда	epravda.com.ua	5,735	2.9M	1,678
babel	Бабель	babel.ua	2,302	1.1M	1,459

Fields

column	type	description
id	string	sha1 of the URL, first 20 hex chars
source	string	outlet key, as in the table above
url	string	article URL as fetched
canonical_url	string	canonical URL declared by the page
hostname	string	host the article was served from
date	string	publication date, YYYY-MM-DD — the partition key
title	string	headline
author	string	byline where the page declares one (89% of rows)
description	string	page description / standfirst
text	string	extracted body text, plain
markdown	string	same body with headings and lists preserved
categories, tags	string	comma-joined, as declared by the page (85% of rows)
image	string	lead image URL
sitename	string	site name declared by the page
fingerprint	string	simhash of the body — near-duplicate detection
text_len, n_paragraphs	int32	body length in characters, paragraph count
lang	string	detected language; always uk in this release
html_len	int32	length of the source HTML the text came from
fetched_at	string	when we fetched it, ISO 8601
sitemap_url, sitemap_lastmod	string	where the URL was discovered

How it was collected

Per-outlet discovery from sitemaps or dated archive pages, then trafilatura for boilerplate-free extraction. Language is decided per article, not per URL: Ukrainian-only letters (і/ї/є/ґ) run 4–6% of Cyrillic in Ukrainian and ~0% in Russian, so the ratio separates the two cleanly. 46,363 Russian-language articles were fetched and discarded.

Articles are partitioned by publication date, not fetch date.

Known limitations

• Coverage is ~86%. Roughly 500k Ukrainian articles exist in this window across these outlets. The shortfall is the tails of unian and nv, which rate-limited the crawler; it is spread thinly rather than concentrated in a date gap.
• 1,289 rows (0.3%) fall outside the stated window, mostly evergreen pages whose in-page date is wrong. Filter on date if that matters.
• raw is not deduplicated, and it repeats 184 rows outright (identical id and url, ingested twice). Both are handled in deduplicated — see below.
• Roughly a quarter is not hard news — advice columns, horoscopes, recipes, celebrity items. On a hand-checked day, 63% of articles were datable events and 20% were evergreen filler.
• Trailing calls-to-action survive in some articles, concentrated in obozrevatel and slovoidilo. Filterable from markdown.

Deduplication

deduplicated is raw with near-duplicates removed, by the same method and parameters text-dedup applies — chosen so this corpus can be mixed with Goader/kobza without the two halves carrying different duplicate statistics.

step	removed	remaining
raw	—	429,427
exact row repeats (same id, same url)	184	429,243
MinHash near-duplicates	10,039 (2.3%)	419,204

MinHash over word 5-grams, 250 permutations, LSH banded for Jaccard ≥ 0.7, clusters closed by union-find over band collisions, one document kept per cluster. Band collisions are taken as cluster membership without verifying exact Jaccard, the same approximation text-dedup makes. Within a cluster the earliest published document survives, so what remains is the original report rather than a later aggregator's copy.

Cluster sizes are dominated by pairs — 7,114 of them, then 774 triples — with a thin tail up to one cluster of 121. The large clusters are serial formats rather than syndication: daily casualty reports (Втрати ворога станом на …), fuel-price roundups, exchange-rate posts. Ukrainian outlets mostly rewrite rather than republish, which is why the figure is 2.3% and not the 10–20% typical of a web crawl.

The fingerprint column (trafilatura simhash) is still there if you want to deduplicate raw differently; note that exact fingerprint equality finds only 1.4%, well short of what MinHash finds.

train-mixin

deduplicated, restricted to the stated window, minus every article that asserts one of 42 held-out facts.

It exists for a continual-pretraining experiment that measures how much of this corpus a model actually absorbs. That measurement needs a floor: a set of facts present in the corpus but withheld from training, so that whatever a trained model scores on them is guessability rather than exposure. One article carrying a held-out fact into the stream turns the floor into a tautology, so those articles are removed — 31,452 of them, 7.5%.

in-window deduplicated	417,929
withheld (assert a held-out fact)	31,452 (7.5%)
train-mixin	386,477

Exclusion is at the document level: the article contains a form of the withheld answer and a content word of its description, anywhere in the text. Verified to zero — no article in train-mixin asserts a held-out fact under that test.

Nothing else is filtered. No quality scoring, no boilerplate stripping, no outlet weighting; the stream is meant to look like the news a production recipe would actually use.

It is bound to one version of the probe set (sha256 4fafcf7d…, 128 items, 42 held out). If the item set changes, this config has to be rebuilt. Use deduplicated for anything that is not this experiment.

Loading

from datasets import load_dataset

ds = load_dataset("Goader/ukrainian-news-2026", split="train")          # deduplicated
raw = load_dataset("Goader/ukrainian-news-2026", "raw", split="train")  # original
mix = load_dataset("Goader/ukrainian-news-2026", "train-mixin", split="train")

# or straight from Parquet
import pyarrow.parquet as pq
t = pq.read_table("deduplicated/ukr-news-2026-dedup.parquet",
                  columns=["source", "date", "title", "text"])

Rows are written in publication-date order, so date-range filters skip row groups.

Disclaimer

This dataset is provided as-is, for research. The packagers assert no ownership of the article text, grant no rights over it, and accept no liability for how it is used. If you are a rights holder and want your content removed, open a discussion on this dataset and we will act on it.