RuWiki-Good-Thematized-2026 — Russian Wikipedia quality articles with subjects

This Russian-language collection contains 17,251 Wikipedia articles with at least one subject label from a fixed taxonomy of 28 subjects. It is selected from the 17,413 articles in Russian Wikipedia quality articles, based on the 20261001 dump. Each article retains the original title, readable text, wikitext, direct categories, community quality status and revision links, and adds a topics list. The topics configuration supplies the complete subject definitions. The full category catalog remains in the source collection.

Files at a glance

File	Purpose	Bytes
data/articles.jsonl.gz	Source articles with retained subject labels.	413,342,335
data/topics.jsonl.gz	The 28 subject definitions and their boundaries.	2,671
data/labels.json	Article-ID to subject-slug mapping for replay and analysis.	649,823
data/topic-provenance.jsonl.gz	Frozen initial/final assignments and review status for all Full IDs.	127,895
data/noow.csv	Whole-document text and ordered normalized VW features.	1,118,455,706
data/bow.csv	The same documents with counted VW features.	875,143,956
data/splits.json	Optional fit/calibration/validation memberships by article ID.	291,038

SHA256SUMS covers this repository's published data. Preparation code, environment locks and source/output pins are maintained once in the Full collection.

Selection and subject labels

The source includes namespace-0 articles with active recommended (Добротные), good (Хорошие) or featured (Избранные) community status. Redirects and disambiguation pages are excluded. Multiple active statuses are retained; former-status categories are recorded separately. These are editorial designations at the snapshot date, not independent factual checks.

The taxonomy was inspired by the broad subject domains of PostNauka, SciNews, and BBC News, then adapted to Wikipedia articles and their editorial rubrics. Its 28 subjects are not an identical copy of any of those sources' classes; the published topic definitions specify the boundaries used here.

Subject assignment combines mappings from specific direct editorial rubrics with AI-assisted review of ambiguous rubrics using direct subject categories and, where needed, excerpts from the saved readable text. Mixed editorial rubrics identify possible subjects; they do not assign all those subjects simultaneously. Category ancestry is not propagated. Category mappings are assessed against the subject definitions: a category name being present does not by itself establish a narrow subject, and an interdisciplinary regional-studies category does not automatically establish language scholarship. Incidental mentions, birthplaces, awards, party membership or education alone do not establish a subject. Labels already supported by specific rubrics are retained unless there is explicit contrary evidence.

Labels are partial and were curated with AI assistance; they are not exhaustive annotation or a human gold standard. Category interpretation may be wrong, and relevant subjects may have been missed. Text review was limited to selected passages where categories needed clarification; articles were not read in full. A missing label is not evidence that the article lacks that subject. Multiple labels are allowed when each has independent support. They can arise from several specific independent rubrics, as well as from additional review; a mixed rubric alone never establishes several labels. Articles are included when their subject list is nonempty: 162 source articles have no assigned subject label and are omitted. No coverage quota, topic balancing or maximum label count was used. Subject frequencies and intersections are consequently uneven.

Of the 17,413 source articles, 14,945 required no additional review: 13,984 initially had one subject and 961 had several. The remaining 2,468 articles had evidence from 44 ambiguous editorial rubrics and received additional AI-assisted review. Of these, 2,306 were retained and 162 were excluded because their final subject lists were empty. In total, 2,054 articles had their subject lists changed. The final Thematized collection contains 17,251 articles: 15,599 with one subject and 1,652 with several. Additional review records a step in annotation: it does not imply an automatic label change and is not a confidence score. Articles without additional review are not thereby certified correct.

The original JSON records have no NLP preprocessing or model scoring. The optional preparation workflow below produces separate NLP CSV files without a research train/test split; a fixed reference partition is supplied separately in data/splits.json. The corpus split is a storage partition. Source text has not been shortened or rewritten for subject selection. No exact duplicate source text was removed. This collection is not a representative sample of Russian writing.

Quick start: load the published files

The articles configuration is the default. articles and topics use compressed JSON Lines and the corpus split, with no dataset script.

from datasets import load_dataset

articles = load_dataset(
    "TopicNet/RuWiki-Good-Thematized-2026", "articles", split="corpus"
)
topics = load_dataset(
    "TopicNet/RuWiki-Good-Thematized-2026", "topics", split="corpus"
)
print(len(articles))  # 17251
print(len(topics))    # 28

From a downloaded repository, load the JSON Lines files directly:

articles = load_dataset(
    "json", data_files={"corpus": "data/articles.jsonl.gz"}, split="corpus"
)
topics = load_dataset(
    "json", data_files={"corpus": "data/topics.jsonl.gz"}, split="corpus"
)

The prepared NLP files data/noow.csv and data/bow.csv contain the same 17,251 article IDs, in source order, with string columns id, raw_text, vw_text. Use the CSV reader for these files, separately from the repository's JSON Lines configurations:

noow = load_dataset(
    "csv",
    data_files={"corpus": "hf://datasets/TopicNet/RuWiki-Good-Thematized-2026/data/noow.csv"},
    split="corpus",
)
bow = load_dataset(
    "csv",
    data_files={"corpus": "hf://datasets/TopicNet/RuWiki-Good-Thematized-2026/data/bow.csv"},
    split="corpus",
)

From the root of the downloaded repository:

noow = load_dataset(
    "csv", data_files={"corpus": "data/noow.csv"}, split="corpus"
)
bow = load_dataset(
    "csv", data_files={"corpus": "data/bow.csv"}, split="corpus"
)

CSV is the primary prepared format. The explicit "csv" builder above is required: these files are not additional YAML configurations, because combining JSON and CSV builders in one repository configuration did not load reliably. The published CSV files have been fully read with Hugging Face Datasets.

Optional reference splits

data/splits.json maps each partition name to an ordered list of article IDs. The lists are disjoint and together cover all 17,251 articles:

Partition	Documents	Intended use
fit	10,342	Fit models and training-dependent vocabularies or statistics.
calibration	3,434	Tune parameters and make development decisions.
validation	3,475	Evaluate the selected procedure on held-out documents.

The partition uses multilabel group stratification with approximate proportions 60/20/20 and seeds 20260824 and 20260825 for the two successive splits. Near-duplicate articles, identified by Jaccard similarity ≥0.9 between sets of five-word shingles, stay in the same partition. The procedure found one such pair (article IDs ruwiki/10869327 and ruwiki/10872523 — whose titles are "Город Казань (шмак)" and "Северная звезда (шмак)", respectively); both articles are in fit. No articles were removed.

The same IDs occur in articles, noow.csv, bow.csv and labels.json. The Hugging Face split corpus still contains the whole collection; use the stored memberships when you want this reference partition. After loading articles above, run from the root of a downloaded repository:

import json
from pathlib import Path
from datasets import DatasetDict

splits = json.loads(Path("data/splits.json").read_text(encoding="utf-8"))
index = {doc_id: i for i, doc_id in enumerate(articles["id"])}
partitioned = DatasetDict({
    name: articles.select([index[doc_id] for doc_id in ids])
    for name, ids in splits.items()
})
print({name: len(part) for name, part in partitioned.items()})
# {'calibration': 3434, 'fit': 10342, 'validation': 3475}

For either CSV representation, replace articles with the loaded noow or bow dataset. Selection preserves the order recorded in splits.json. The shared corpus/NLP preparation workflow does not generate this supplementary partition. Its published ngrams use all article bodies, so applying the memberships does not make those features training-only; see the NLP preparation section below. Text-only modeling can use @word and fit any further vocabulary or statistics on fit alone.

To export one VW document per line without loading the CSV into memory, run this from the downloaded repository. Choose an output path outside the repository:

import csv
from pathlib import Path

csv.field_size_limit(2**31 - 1)
with Path("data/noow.csv").open(encoding="utf-8", newline="") as source:
    with Path("../noow.vw").open("w", encoding="utf-8", newline="") as output:
        for row in csv.DictReader(source):
            output.write(row["vw_text"] + "\n")

The field-size setting is needed only for Python's standard csv reader; it does not configure Hugging Face Datasets. Use bow.csv in the same example for counted features. CSV records contain whole documents and may include embedded newlines in raw_text, so splitting the CSV by physical lines is unsafe. The repository does not store a second permanent copy as VW text files.

The NLP preparation section describes the normalized namespaces and text fields.

The sidecar data/labels.json is a JSON object mapping every Thematized article ID to its list of subject slugs, matching both article topics and the CSV @topic namespace. It is not a separate dataset configuration.

import json
from pathlib import Path

labels = json.loads(Path("data/labels.json").read_text(encoding="utf-8"))

Frozen topic provenance

data/topic-provenance.jsonl.gz contains one record for each of the 17,413 Full article IDs, including the 162 articles excluded from Thematized. Records are ordered by numeric page ID. This is frozen annotation metadata, separate from corpus reconstruction and NLP build outputs; preprocessing does not re-annotate it or create new subject evidence. It records the initial specific-rubric assignments and the final accepted assignments, without publishing internal annotation decision logs.

Field	JSON type	Meaning
id	string	Full article ID, ruwiki/<page_id>; join to Full to inspect excluded articles.
initial_topics	array of strings	Subjects supported by specific direct editorial rubrics before additional review.
topics	array of strings	Final subjects; matches data/labels.json for every Thematized ID. An empty list identifies an excluded Full article.
additional_review	boolean	Whether baseline ambiguous-rubric evidence led to additional AI-assisted review; not whether labels changed or a confidence score.
unresolved_rubrics	array of strings	Ambiguous direct editorial rubrics at the initial stage; retained as historical evidence even after review, not a claim that the final subjects remain unresolved.

To stream only reviewed articles retained in Thematized, run this from the root of the downloaded repository. Use not row["topics"] instead of row["topics"] to inspect the excluded articles, joining their IDs to Full:

import gzip
import json

with gzip.open("data/topic-provenance.jsonl.gz", "rt", encoding="utf-8") as source:
    for line in source:
        row = json.loads(line)
        if row["additional_review"] and row["topics"]:
            print(row["id"], row["initial_topics"], row["topics"])

Article fields

Field	JSON type	Meaning
id	string	Stable page identity, ruwiki/<page_id>.
page_id	integer	Russian Wikipedia page ID.
title	string	Article title with spaces, separate from the text.
wikitext	string	XML-decoded revision wikitext.
raw_text	string	Readable text extracted from wikitext; not rendered HTML.
wikitext_sha256	string	SHA-256 of UTF-8 encoded wikitext.
raw_text_sha256	string	SHA-256 of UTF-8 encoded raw_text.
quality_statuses	array of strings	Active recommended, good and/or featured statuses from SQL.
former_quality_categories	array of strings	Direct former-status category names from SQL; may be empty.
categories	array of strings	All direct SQL category names, including maintenance categories.
wikidata_id	string or null	Wikidata QID from SQL page properties, when available.
revision_id	integer	Content revision ID from XML.
revision_timestamp	string	XML revision timestamp in UTC, such as 2026-09-26T22:17:57Z.
metadata_revision_id	integer	SQL page revision ID; may differ from revision_id.
page_url	string	Page-ID URL for the source page and access to its history.
revision_url	string	Permanent URL for the content revision.
topics	array of strings	Sorted unique subject slugs; at least one in every retained article.

The source fields are preserved exactly; only topics is added. The table describes the stored JSON types. Hugging Face Datasets 5.0.0 infers revision_timestamp as timestamp[s] and former_quality_categories as List(null) for this release, whose former-status lists are empty. Nullable scalar fields remain nullable when loaded.

For example, the following is an actual record projected onto a few fields (the text and remaining fields are omitted here):

{
  "id": "ruwiki/21",
  "page_id": 21,
  "title": "Социология",
  "quality_statuses": [
    "featured"
  ],
  "wikidata_id": "Q21201",
  "revision_id": 155304860,
  "metadata_revision_id": 155304860,
  "revision_timestamp": "2026-09-26T22:17:57Z",
  "page_url": "https://ru.wikipedia.org/?curid=21",
  "revision_url": "https://ru.wikipedia.org/w/index.php?oldid=155304860",
  "topics": [
    "social-sciences-education"
  ]
}

Topic fields and distribution

Each topic record contains the definition used to assign that subject. Join article topics entries to topic slug.

Field	JSON type	Meaning
slug	string	Stable subject identifier used in article topics.
name_en	string	English subject name.
name_ru	string	Russian subject name.
definition_and_boundaries	string	Russian definition and boundaries between subjects.

Counts overlap because an article may have several subjects.

Subject	Slug	Articles
Mathematics and statistics	mathematics	158
Physics	physics	220
Chemistry	chemistry	132
Astronomy and space	astronomy-space	283
Earth science	earth-science	63
Geography and places	geography	767
Biology	biology	1,370
Medicine and health	health	207
Environment and agriculture	environment-agriculture	87
History and archaeology	history-archaeology	4,177
Military affairs and security	military-security	2,332
Politics and law	politics-law	1,140
Economics and business	economy-business	254
Philosophy	philosophy	108
Religion and mythology	religion-mythology	814
Language and linguistics	language	264
Literature	literature	1,132
Visual and performing arts	visual-performing-arts	755
Architecture and structures	architecture	428
Film, television and animation	film-television-animation	850
Music	music	952
Games	games	784
Sport	sport	754
Media and communication	media-communication	123
Food and gastronomy	food	92
Technology and industry	technology	356
Transport	transport	376
Society and education	social-sciences-education	254

Subjects per article	Articles
1	15,599
2	1,400
3	197
4	38
5	13
6	3
7	1

Sources and revision semantics

The source is the Russian Wikipedia 20261001 dump. The source and output pins record the names, byte sizes and SHA-256 digests of all six inputs: page, page properties, category links and link targets SQL dumps, plus the multistream article XML and its index. Dump mirrors may later remove this dated snapshot; the pins identify the inputs independently of their availability.

SQL provides selection, titles, page properties and category relationships. XML provides article content, revision_id and revision_timestamp. In the 17,413-article source collection, 11 articles have differing SQL and XML revision IDs; the metadata and content should not be assumed to describe one atomic revision. Use revision_url to identify the exact text revision, and retain both IDs when recording provenance. Live pages and categories can differ from this snapshot.

Readable extraction uses mwparserfromhell 0.7.2 and deterministic markup cleanup. It does not render Wikipedia or expand templates fully. Templates, infoboxes, tables, references, captions, formulae and formatting may be removed or simplified; fragments of markup or awkward text may remain. wikitext is retained so readers can inspect these losses or apply a different extraction procedure. Neither raw_text nor the category metadata is a manually corrected annotation.

NLP preparation

The standalone preparation code and full instructions are maintained in Full. Run the preparation commands from Full's preprocessing/ directory, including when preparing only Thematized. Local repository directories may have any names. Replace the paths below with your own absolute paths: work_dir is a separate working directory, and selected_data points to this collection's data/ directory. Choose fresh output and scratch paths on a disk with enough space; corpus output and scratch inside a published clone are rejected. Build and verify require Python 3.11.15 because VW escaping depends on its Unicode database; verify needs no NLP packages. On Linux x86-64, install the supplied locked environment. requirements.txt includes mwparserfromhell==0.7.2 for corpus extraction.

cd "/absolute/path/to/full/preprocessing"
work_dir="/absolute/path/to/ruwiki-rebuild"
selected_data="/absolute/path/to/selected/data"

conda create --name ruwiki-nlp --file conda-linux-64.lock.txt
conda activate ruwiki-nlp
python -m pip install -r pip-linux-x86_64.lock.txt
python prepare.py --stage nlp --action download --dataset selected \
  --sources-dir "$work_dir/nlp-sources"
python prepare.py --stage nlp --action build --dataset selected \
  --sources-dir "$work_dir/nlp-sources" \
  --output-dir "$work_dir/nlp-output" \
  --scratch-dir "$work_dir/nlp-scratch"
python prepare.py --stage nlp --action verify --dataset selected \
  --sources-dir "$work_dir/nlp-sources" \
  --output-dir "$work_dir/nlp-output"

Local paths above are independent of Hugging Face repository names. Online loading examples and cross-repository links use the published Hugging Face IDs. If those IDs change, the links and the downloader's pinned source addresses need to be reviewed separately; renaming a local directory does not change them.

--stage nlp is the default; commands without it continue to work. Use --dataset both to prepare both collections. Download checks pinned source snapshots and supplies the pinned NLTK stopwords archive. Build runs offline. Keep output and scratch separate from published data/ and use fresh directories for a new build. Full's instructions explain reuse of local JSONL files and access to restricted source repositories.

After the structural NLP checks, the shared entry point independently compares CSV and Thematized label byte sizes and SHA-256 hashes with published expectations in the corpus contract, reporting PUBLISHED_NLP_OK. It does not rely on the new build manifest as its authority and does not replay NLP normalization.

The workflow produces $work_dir/nlp-output/selected/noow.csv and bow.csv with exactly id, raw_text, vw_text columns for all 17,251 selected IDs, preserving article order. CSV raw_text is the title, two LF characters and the original readable body with its leading CR/LF characters removed. Original JSON raw_text remains the readable body alone; the source JSON files are unchanged.

The normalized VW namespaces are @title, @word, @bigram, @trigram, @category, then @topic. @word contains the body only. @category retains all direct categories, including service and maintenance categories. @topic contains the accepted subject slugs described above, not raw Wikipedia rubric names. labels.json maps article IDs to those same subject slugs. These labels remain partial and AI-assisted; preprocessing does not improve their validity. NOOW preserves normalized token order within each namespace, not raw prose. BOW counts precisely the same tokens separately in every namespace.

Normalization uses spaCy Russian tokenization, pymorphy3 lemmas and the union of spaCy and pinned NLTK Russian stopwords. Punctuation is removed; tags, URLs and email addresses become placeholders. Cyrillic stress marks are removed and lexical Cyrillic hyphenated compounds are retained with ASCII hyphens. Names are not masked. There is no article removal, token trimming, research split, model training or chain extraction. Build prints phase and document-count progress; the manifest reports empty @word articles by count and ordered IDs without dropping them.

Bigrams and trigrams use independent NLTK PMI selection after frequency ≥5, with nbest(32100)[100:] for each order. This selection uses all Thematized article bodies, independently of Full; it is corpus-global rather than fitted only on a future training partition. For text-only experiments, use only @word to avoid exposing category, title or subject metadata to the model. The workflow does not establish an evaluation without information from held-out documents.

Load a generated CSV directly:

from pathlib import Path

# Use the same work directory as in the shell commands.
work_dir = Path("/absolute/path/to/ruwiki-rebuild")
output_dir = work_dir / "nlp-output/selected"
noow = load_dataset(
    "csv", data_files={"corpus": str(output_dir / "noow.csv")}, split="corpus"
)
bow = load_dataset(
    "csv", data_files={"corpus": str(output_dir / "bow.csv")}, split="corpus"
)

A measured NLP build of both Full and Thematized from ready JSONL snapshots took 42 min 5.83 s, with 10.58 GiB peak resident memory, approximately 3.74 GiB of final output and 0.53 GiB of scratch. These are measurements from one environment, not Thematized-only estimates or measurements of replay from raw Wikimedia dumps.

Rebuild Thematized from source data

Full's corpus replay instructions cover the six pinned 20261001 dumps, explicit download resumption with --resume, offline construction and independent published-payload verification. The shared source and output contract contains source dump and annotation pins. The new public path from raw dumps has not yet been run end to end; its runtime and resource requirements remain unmeasured.

For the complete path, keep the same Full preprocessing/ working directory and path variables as in the NLP commands above:

python prepare.py --stage corpus --action download --dataset both \
  --sources-dir "$work_dir/dump-sources"
# Add --resume explicitly when continuing an interrupted download.
python prepare.py --stage corpus --action build --dataset both \
  --sources-dir "$work_dir/dump-sources" \
  --annotations-dir "$selected_data" \
  --output-dir "$work_dir/corpus-output" \
  --scratch-dir "$work_dir/corpus-scratch"
python prepare.py --stage corpus --action verify --dataset both \
  --sources-dir "$work_dir/dump-sources" \
  --output-dir "$work_dir/corpus-output"

This writes $work_dir/corpus-output/full/articles.jsonl.gz and categories.jsonl.gz, plus $work_dir/corpus-output/selected/articles.jsonl.gz and topics.jsonl.gz. The annotation directory must already contain this collection's labels.json and topics.jsonl.gz. Reconstruction retains the published mapping and does not rerun AI annotation or establish new evidence for label correctness.

If a previously accepted Full corpus is already available, build only Thematized without rereading the dumps. From the same Full preprocessing/ directory, resolve Full's data path:

full_data="$(cd ../data && pwd -P)"
python prepare.py --stage corpus --action build --dataset selected \
  --sources-dir "$full_data" \
  --full-dir "$full_data" \
  --annotations-dir "$selected_data" \
  --output-dir "$work_dir/selected-output" \
  --scratch-dir "$work_dir/selected-scratch"
python prepare.py --stage corpus --action verify --dataset selected \
  --sources-dir "$full_data" \
  --output-dir "$work_dir/selected-output"

This Thematized-only workflow has been verified byte for byte against the published archive of all 17,251 articles, the labels file and the topic definitions.

Use fresh output and scratch directories. Corpus verification checks exact compressed output bytes against the separately bundled published payload pins; it does not treat a newly generated manifest as independent evidence. To prepare NLP CSVs from $work_dir/corpus-output, run the earlier NLP commands with --sources-dir "$work_dir/corpus-output" and fresh --output-dir "$work_dir/replay-nlp-output" and --scratch-dir "$work_dir/replay-nlp-scratch". Run NLP download first to supply NLTK stopwords, then offline build and verify.

Preparation code is MIT licensed in the source collection. Article text and its normalized derivatives retain the CC BY-SA 4.0 terms and attribution obligations below.

Attribution, license and use

Article text and wikitext are attributed to Russian Wikipedia contributors and distributed under Creative Commons Attribution-ShareAlike 4.0. Each record's revision_url identifies the revision used; page_url provides access to the page and its edit history for contributor attribution. Follow attribution and share-alike requirements when redistributing text or adaptations, identify your changes, and preserve appropriate source links. Wikipedia can contain quotations or other third-party material with additional terms. Wikidata identifiers refer to Wikidata, whose structured data is available under CC0; their presence does not change the license of the article text. Images and other media binaries are not included.

Potential uses include Russian text analysis, inspection of markup extraction and research on Wikipedia editorial metadata or category relationships. This is a small, editorially selected portion of Russian Wikipedia, not a representative sample of Russian writing or an evenly balanced subject collection. Quality status does not guarantee accuracy, neutrality, completeness or suitability for any downstream application. Articles can contain sensitive, biased, outdated or offensive material. The partial subject labels described here do not establish a complete reference annotation. The optional reference splits above provide one document partition; further downstream preprocessing, deduplication and any alternative split policy must be specified for each study.

File verification

Run from this downloaded repository's root:

sha256sum -c SHA256SUMS

SHA256SUMS covers all seven data files, excluding this README and Git/service metadata. Full has its own manifest for its data and shared preparation files. The shared corpus contract independently pins expected published reconstruction payloads. A successful local checksum check establishes agreement with the supplied manifest; use a trusted repository revision to establish provenance.