fastpixels/odia-wikipedia-sep26

Odia Wikipedia, cleaned for LLM training

Every article on Odia Wikipedia in the 2026-09-01 dump (orwiki-20260901-pages-articles.xml.bz2, the latest complete dump when built on 2026-10-02), as clean GitHub-flavoured Markdown, one article per record: 18,683 articles, 4,514,702 Odia words, 87 MB of UTF-8 text.

Download

Two files are ready to take and use as they are:

• orwiki-20260901-trainingready.jsonl.gz: the corpus. 18,683 Odia Wikipedia articles as clean Markdown, one JSON object per line, 4,514,702 Odia words (17 MB gzipped, 87 MB unpacked). This is the file to train on.
• translations/english-odia-pairs.jsonl: English-to-Odia translation pairs, if you want them separately, ready for training: 726 pairs, one per line, {"english": …, "odia": …}. They are English paragraphs and headings found in these articles, with their Odia translations, which the corpus has in place of the English.

import gzip, json
articles = [json.loads(line) for line in gzip.open("orwiki-20260901-trainingready.jsonl.gz", "rt", encoding="utf-8")]
pairs = [json.loads(line) for line in open("translations/english-odia-pairs.jsonl", encoding="utf-8")]

The pairs come from the translation table, translations/english-to-odia.jsonl, which also keeps each translation's article, kind, checks and notes. Left out of the pairs: junk, or names and titles kept in English (85); the English side has Odia in it (39); failed an automatic check (25); template residue in the English (1).

Everything else in this repository is how they were made, and what it takes to make them again.

What is here

File	What it is
orwiki-20260901-trainingready.jsonl.gz	the training-ready corpus, one JSON object per line (fields below), gzipped; build also writes it unpacked, as orwiki-20260901-trainingready.jsonl
translations/english-odia-pairs.jsonl	English-Odia translation pairs for training, english and odia only
translations/english-to-odia.jsonl	the translation table the pairs come from: each translation with its article, kind, checks and notes
excluded.jsonl	every page of the dump that is not in the corpus: id, revid, title, reason, detail
removed-blocks.jsonl	blocks taken out of articles: citations, junk, prose awaiting translation
orwiki-20260901-trainingready-build.json	build statistics (counts per exclusion reason; the pages are in excluded.jsonl)
annotations/*.jsonl	topics, translation flags, Sarvam-1 scores and the review queue, joined on id
pipeline.py	one command that rebuilds everything from the cached inputs, in order, and checks it
prepare.py, check.py, translate.py, annotate.py, score_bpb.py	the steps
METHODOLOGY.md	every step and rule applied to the data, with the evidence and counts
LEARNINGS.md	what building this corpus taught us, and ideas for next steps
curation/junk-paragraphs.jsonl	paragraphs judged by hand not to be content, with the reason; build drops them
reviews/reviews.jsonl	review decisions (keep, drop, fix, paragraphs to drop); build applies them
odia_text.py	the Odia text rules the steps share: normalisation, Odia words, digits
edaapp/	the review web app: browse the dataset, read the methodology, record review decisions
publish_hub.py	uploads the dataset to Hugging Face: the files git tracks, without edaapp/
LICENSE	CC BY-SA 4.0, for the data and the code
raw/	rebuild inputs: article index, dump provenance, rendered HTML, annotation inputs, model scores

All outputs are JSON, JSON lines or Markdown, to read with any editor or jq. Every file here is tracked in git, all under 50 MB and none through Git LFS: the corpus as orwiki-20260901-trainingready.jsonl.gz, while the unpacked orwiki-20260901-trainingready.jsonl is left out and rebuilt by pipeline.py from the inputs in raw/, translations/, curation/ and reviews/.

Record format

Field	Meaning
id	page id on or.wikipedia.org
title	article title
url	article URL
revid	revision in the dump; https://or.wikipedia.org/w/index.php?oldid=<revid> is exactly this text
timestamp	when that revision was saved
text	the article (format below)
words	Odia words in text: runs of Odia letters, title included; numbers and Latin words don't count
chars	characters in text (Unicode code points), title line and Markdown included
odia_ratio	share of non-space characters in the Odia block (odia_text.odia_ratio)
tables	data tables in text, as Markdown tables
translated_paragraphs	paragraphs and headings machine-translated from English (0 = all native Odia)
bot_created	the page carries {{ବଟ୍ ତିଆରି}}, made by a bot (in 2026-09 only year pages, all excluded)
stub	the page carries a stub template ({{ମୁଣ୍ଡିଆ}}, {{ଅଧାଗଢ଼ା}})
templated_share	share of the article's Odia words in template sentences (found in 5+ articles); high = formulaic

text is GitHub-flavoured Markdown: # title, the lead, ##/### section headings, paragraphs separated by a blank line, - / 1. lists (sublists indented to their parent's content column), data tables as GFM tables (| a | b |), and math as $...$ ($$...$$ on its own line). Links and emphasis are reduced to their text.

It is checked as Markdown, not just shaped like it. Every article parses (markdown-it, CommonMark + GFM tables) into exactly the headings, list items and tables it is meant to have, with no accidental emphasis, links, code, HTML, block quotes or rules. pandoc's GFM reader agrees on a random sample of 405 articles. Text from the page that would read as markup is backslash-escaped: \*, \_, \$, \<alt>, 1\. or \- at the start of a line. There are 2,189 escapes in 519 articles, about 63 per million characters.

A short example record:

{
  "id": 3008,
  "title": "କର୍ଣ୍ଣାଟକ",
  "url": "https://or.wikipedia.org/wiki/%E0%AC%95%E0%AC%B0%E0%AD%8D%E0%AC%A3%E0%AD%8D%E0%AC%A3%E0%AC%BE%E0%AC%9F%E0%AC%95",
  "revid": 498992,
  "timestamp": "2023-08-27T17:20:28Z",
  "words": 77,
  "chars": 576,
  "odia_ratio": 0.9052,
  "tables": 0,
  "translated_paragraphs": 0,
  "bot_created": false,
  "stub": true,
  "templated_share": 0.0,
  "text": "# କର୍ଣ୍ଣାଟକ\n\nକର୍ଣ୍ଣାଟକ (କନ୍ନଡ: ಕರ್ನಾಟಕ) ଦକ୍ଷିଣ-ପଶ୍ଚିମ ଭାରତର ଏକ ରାଜ୍ୟ । 1956 ମସିହା ନଭେମ୍ବର 1 ତାରିଖରେ ରାଜ୍ୟ ପୁନର୍ଗଠନ ଆଇନ ବଳରେ ଏହି ରାଜ୍ୟ ସ୍ଥାପିତ ହୋଇଥିଲା । ଏହାର ପୂର୍ବ ନାମ ମହୀଶୂର ଥିଲା । 1973 ମସିହାରେ ଏହାର ନୂତନ ନାମକରଣ ହେଲା କର୍ଣ୍ଣାଟକ । କର୍ଣ୍ଣାଟକର ପଶ୍ଚିମରେ ଆରବ ସାଗର, ଉତ୍ତର-ପଶ୍ଚିମରେ ଗୋଆ, ଉତ୍ତରରେ ମହାରାଷ୍ଟ୍ର, ଉତ୍ତର-ପୂର୍ବରେ ତେଲେଙ୍ଗାନା, ପୂର୍ବରେ ଆନ୍ଧ୍ର ପ୍ରଦେଶ, ଦକ୍ଷିଣ-ପୂର୍ବରେ ତାମିଲନାଡ଼ୁ ଓ ଦକ୍ଷିଣ-ପଶ୍ଚିମରେ କେରଳ ଅବସ୍ଥିତ । ରାଜ୍ୟର ମୋଟ ଆୟତନ 1,91,791 ବର୍ଗକିଲୋମିଟର ଓ ଭାରତର ମୋଟ ଆୟତନର 5.83 ପ୍ରତିଶତ । ଏହି ରାଜ୍ୟର ଜିଲ୍ଲା ସଂଖ୍ୟା 30 ଓ ସରକାରୀ ଭାଷା କନ୍ନଡ଼ ।\n\n## ଭାଷା ଓ ସାହିତ୍ୟ\n\nକନ୍ନଡ଼ ହେଉଛି ଏହାର ମାତୃଭାଷା ।"
}

How it was made

1. Download. orwiki-20260901-pages-articles.xml.bz2 from dumps.wikimedia.org, with its SHA-1 (4ace67b6c5412a78565e463066f9450bfff6df6b) checked against the dump's dumpstatus.json. The build needs only each article's id, title, revision id, timestamp and bot/stub flags, so the dump is reduced to that index (raw/orwiki-20260901-articles.jsonl, with its provenance in raw/orwiki-20260901-dump.json) and deleted; download fetches it again.
2. Render. Every main-namespace page that is not a redirect (21,095 pages) was fetched as Wikipedia's own rendering (Parsoid HTML) of the exact revision in the dump, from /w/rest.php/v1/revision/<revid>/html. Odia articles build whole sentences out of templates: '''{{PAGENAME}}''' ଏକ ଭାରତୀୟ {{TownType|M}} in over 1,600 town stubs, and {{Birth date}}, {{convert}}, {{flag}} and other Lua-module templates across the wiki. Stripping the wikitext would leave holes in those sentences. The rendered HTML has what a reader sees.
3. Clean (HTML to text). Prose, headings, lists and math are kept. Dropped:
   • citations, reference lists, and reference-type sections (ଆଧାର, ଟୀକା, ଆହୁରି ଦେଖନ୍ତୁ, ବାହାର ଲିଙ୍କ, ଅଧିକ ପଢ଼ନ୍ତୁ, ଗ୍ୟାଲେରୀ, … and their English equivalents)
   • infoboxes, navboxes, layout tables, images, galleries, captions, maps. Data tables (wikitable) are kept as Markdown tables. A cell spanning rows repeats in each row, a cell spanning columns fills the first one, and image-only columns are removed. Tables hold 169,433 Odia words (3.8%) in 1,721 articles: lists of districts, constituencies, award winners and office holders. For prose only, drop the lines starting with |.
   • hatnotes, maintenance and stub banners, coordinates, pronunciation (IPA), sister-project boxes, archive notes ("Archived … at the Wayback Machine"), template error messages
   • list items that are only an external link or a book citation, under any heading
   • hand-typed reference markers ([2], [୪]) and broken wikitext that renders as text ([[, {{, '', stray File:…|thumb| lines)
   • bare URLs typed into the prose, and raw HTML pasted into the wikitext (a Content Translation bug) that renders as text
   • empty sections, and parentheses emptied by the removed pronunciations
   • external-link lines written in Odia (IMDb, Facebook, Instagram, Twitter, official website, …), dropped by template name
   • conversion leftovers found by the review queue: Content Translation <a href=… cx-link> tags, raw {| … |} wikitable text, template parameters shown as text (Quote box|width=…), HTML attributes, Category: text, and the entities & / 
 (69 fixes)
   • wikitext headings typed mid-line (… । == ଇତିହାସ ==) become real headings (19)
   • numeric superscripts are written as LaTeX math, like the rest of the math: 10<sup>26</sup> becomes $10^{26}$ and km<sup>2</sup> becomes km$^{2}$ (283). Flattened, they would read 1026 and km2. Powers of ten typed without the superscript upstream, 6.1 x 108, become $6.1 \times 10^{8}$ (13). Short subscripts too: H<sub>2</sub>O becomes H$_2$O (303); chemistry markup is $\ce{…}$
   • paragraphs judged by hand not to be content: test edits, colour legends of tables whose colours are gone, a leaked timeline template, pasted search-result snippets (189; each with its reason in curation/junk-paragraphs.jsonl)
   • paragraphs a reviewer marked fix, replaced by their fixed text: English lists, tables and passages translated into Odia with the names transliterated, and a wrong name corrected (26; each with its source in curation/paragraph-fixes.jsonl)
   • English inside articles (blocks with more than twice as many Latin as Odia letters): citations are removed (104), paragraphs and headings are replaced by their Odia translation (861, see translations/english-to-odia.jsonl and METHODOLOGY.md), and lists, tables and names stay as they are (4,847)
4. Fix known typos. Bots copied some misspellings into hundreds of articles; each fix is tied to its context: ପରୁଷ→ପୁରୁଷ (917).
5. Normalise. normalize_odia from odia_text.py (ୟ written as ଯ + nukta becomes U+0B5F). Odia digits become ASCII (୧୯୪୭ → 1947; 689,041 digits), in text, headings, tables and math, so numbers look the same everywhere, titles included (url keeps the page's real name). A | typed for the danda after Odia text becomes । (and || becomes ॥). Soft hyphens, zero-width spaces, word joiners and BOMs are removed, and runs of spaces are collapsed. ZWJ and ZWNJ stay, because Odia spelling uses them. There is no NFC or other Unicode normalisation (by design).
6. Filter. 2,412 pages were left out, each listed in excluded.jsonl with its id, revision and reason. Boilerplate pages (year pages, date pages without events, empty film-year lists) say nothing beyond their title. Fact-bearing stubs stay, with templated_share for down-weighting.

Reason	Pages
year page	1,864
date page without events	260
under 5 Odia words	137
disambiguation	112
empty list page	19
reviewer: drop	12
duplicate text	7
main page	1

Size

The median article has 151 Odia words (10th percentile 58, 90th 428). 1,006 articles have more than half of their words in template sentences (templated_share > 0.5: bot-made villages, towns and film pages), and 1,910 are marked as stubs. 310 articles (54,816 Odia words) have odia_ratio under 0.6, mostly from English bibliographies and numeric tables; a threshold of 0.6 (the default of odia-llm-trainer's odia-build-cpt --min-odia-ratio) skips them.

Odia words per article	Articles	Words
0–49	1,392	41,522
50–199	10,637	1,274,479
200–999	6,200	2,228,752
1,000–4,999	428	766,474
5,000+	26	203,475

Annotations

Extra fields per article live next to the corpus, one JSON-lines file each (with a .json description), joined on id. A file named *.paragraphs.jsonl has one row per paragraph (id, para). Paragraph i of an article is text.split("\n\n")[i], and paragraph 0 is the # title heading. Annotations that depend on the text carry text_sha1 (sha1 of the scored text), so a rebuild that changes an article makes them visibly stale.

• annotations/bpb.jsonl: Sarvam-1 bits per byte of every article, and a review queue of likely data problems (garbled or wrong-script text, untranslated English, boilerplate, odd tables, conversion leftovers) from paragraph and article bpb extremes. Columns: id, text_sha1, bpb, bpb_text, bpb_pct, bytes, tokens, bits, paragraphs, unscored_paragraphs, extreme_high, extreme_low, extreme_share, review_rank, review_type, review_para, review_reasons. Depends on the text: rows carry text_sha1.
• annotations/bpb.paragraphs.jsonl: Sarvam-1 bits per byte of every paragraph except the title (paragraph i = text.split('\n\n')[i]), with its percentile within its kind and length band. Each row's bits, bytes, tokens, pieces and scoring run are in the score store, joined on para_sha1 (see 'joins'). Columns: id, para, para_sha1, kind, bpb, bpb_group, bpb_pct, bpb_z, flag, latin_share, other_script_share, english_words, markup, verse, repeats, self_repeat, near_copies. More columns, score_run, bytes, tokens, bits, pieces, are read from raw/bpb/scores.jsonl.gz on para_sha1 (kept there once per key). Depends on the text: rows carry para_sha1.
• annotations/topics.jsonl: Subject topics of each article from its categories, its Wikidata item and its title, for weighting a school-oriented training mix. One JSON object per line (topics.jsonl), one per article of the corpus, in corpus order. See quality/topics.md. Columns: id, title, categories, topics, primary_topic, is_person, odisha, school_relevant, topic_source, wikidata, topic_scores.
• annotations/translation.jsonl: Articles created or largely written with Content Translation (CX) or MDWiki's CX-based translation dashboard: machine-assisted translations, mostly from English. One JSON object per line (translation.jsonl), one per article of the corpus, in corpus order. See quality/translation.md. Columns: id, ct_created, ct_any, ct_tags, mdwiki_created, mdwiki_any, tool_bytes_share, translated, translated_at, source_lang, source_title, source_revid, source_section, created, creator_is_bot, revisions, edits_after_creation, bot_edits_after_creation, editors_after_creation, bytes.

Reports: quality/bpb.md, quality/review-first.md, quality/topics.md, quality/translation.md.

Review decisions

Review decisions are kept in reviews/reviews.jsonl, one append-only JSON event per line: ts, dataset (odia-wikipedia), id, title, verdict (keep, drop, fix or null), note, drop_paragraphs (a list of {"para": i, "sha1": …}) and text_sha1 (the text reviewed). They were recorded with edaapp, the review web app in edaapp/ (cd edaapp && uv run edaapp), which writes them to this file; appending events by hand works as well. build applies the latest event per article, so every event carries the article's complete decision: articles marked drop are left out, and dropped paragraphs are removed, matched by the sha1 of their text (not by position), so they survive rebuilds. Paragraph 0 (the title) is never dropped. This build applied 217 reviews: 12 articles dropped, 74 paragraphs dropped, 0 still marked fix, and 0 paragraph decisions whose text is no longer in the article. 26 paragraphs of articles marked fix have been fixed (curation/paragraph-fixes.jsonl, made by translate.py fixes and merge-fixes). Use --no-reviews to build without the decisions.

Using it

import gzip, json
docs = [r["text"] for r in map(json.loads, gzip.open("orwiki-20260901-trainingready.jsonl.gz", "rt", encoding="utf-8"))
        if r["templated_share"] < 0.8]  # e.g. down-weight or skip formulaic stubs

jq -r 'select(.reason == "year page") | .title' excluded.jsonl | head   # why a page is missing
gzip -dc orwiki-20260901-trainingready.jsonl.gz | jq -c 'select(.chars >= 500 and .chars < 600) | {id, title, words, chars}' | head

• In odia-llm-trainer, odia-build-cpt --local orwiki-20260901-trainingready.jsonl --local-upsample 1 adds all of it to a continued-pretraining build. --local upsamples 3× by default, which is meant for textbooks. The builder's own wikipedia source still reads the older Hugging Face snapshot (wikimedia/wikipedia, 20231101.or).
• The text is already normalised with normalize_odia, so a pipeline that applies it again (and line dedup) barely touches it.

Origin

Built from 2026-09-24 to 2026-10-01 inside odia-llm-trainer, a project on Odia language models, as its data/odia-wikipedia/ folder, then moved here with everything needed to rebuild it: the rendered HTML, the annotation inputs, every Sarvam-1 score, the translations, the curated junk paragraphs and the review decisions. edaapp, the web app that recorded those decisions, followed on 2026-10-02. METHODOLOGY.md and LEARNINGS.md keep that project's other names: src/, cpt.py and odia-build-cpt (its training-data builder), the eval harness and experiments (E01, E03, …).

License

Everything in this repository, the data and the code, is licensed CC BY-SA 4.0 (the full text is in LICENSE). The text is by Odia Wikipedia contributors: anything derived from it must keep that license and credit Wikipedia. Each record's revid names its exact source revision, whose history lists the authors.

Rebuild

uv run pipeline.py           # build, annotate, scores, checks, from raw/ (a few minutes)
ODIA_WIKI_CONTACT=you@example.org uv run pipeline.py --fetch   # from scratch (~2 h)

The steps one by one:

uv run prepare.py download   # newest complete dump, or --dump YYYYMMDD
ODIA_WIKI_CONTACT=you@example.org uv run prepare.py render  # ~2 h, resumable
uv run prepare.py build      # about a minute; --min-chars N, --markdown
uv run check.py              # exit 1 if any check fails

render needs contact details in the user-agent (ODIA_WIKI_CONTACT). Wikimedia throttles anonymous bulk clients to about one request a minute per connection. With contact details and 6 parallel requests it ran at about 3 pages/s (21,095 pages in about 2 hours), backing off on the occasional 429 as Retry-After asks. Rendered chunks are cached in raw/html/<date>/, so a rerun fetches only what is missing. The script writes only inside this directory. uv keeps its environment in its own cache.

Publishing

GitHub (https://github.com/odia-genai/odia-wikipedia-sep26) holds this repository with its history. The Hugging Face dataset is the same files without edaapp/, one commit per publish: commit and push first, then uv run publish_hub.py (--dry-run lists what would change). It deletes on the Hub what git no longer tracks and checks every file afterwards.