robbd/wikiparse-2606

wikiparse 2606

Eleven Parquet configs derived from the enwiki-20260601 dump of English Wikipedia: ten metadata tables from the citation and category pipelines plus one article-text tier. The useful unit here is a table, not a monolithic corpus download: each config has one grain, one schema, one checksum, and one job it is reasonably good at.

Coverage

Ten metadata tables (20,637,822 rows in total) plus one article-text tier. Source epoch: the processed enwiki-20260601 corpus (256,564 multistreams). Each count below is the exact row count of the published payload; none is a sample or an estimate.

Config	Rows	What it is good for
refs	256,564	Raw ref and cite syntax counts by multistream member.
articles	5,135,607	Article-level citation counts and parser-quality signals.
categories	529,402	Category-level article and citation aggregates.
publishers	20,000	A bounded publisher-string projection with quality coordinates.
templates	37	Citation-template families by parser outcome.
clusters	1,511,504	Exact-signature citation clusters for deterministic repair work.
streams	256,564	Per-stream custody and citation-spine coverage signals.
disambig-nodes	310,529	Typed disambiguation nodes with explicit relation status.
disambig-links	723,486	Heuristic disambiguation-title candidate links.
redirects	11,894,129	Redirect edges with target-existence and disambiguation flags.

articles-text-p0 publishes the priority-0 article-text delivery slice of the retained-article content tree: 908,270 rows in 1,003 shards (2,552,386,223 payload bytes), or 12.8% of the 7,073,577 retained-article rows in content v1. Rows carry the retained article Markdown, its categories, detector flags, and per-article provenance (stream_id, article_sha256, source archive hash). Composition, computed from the reviewed context index:

• 621,366 rows (68.4%) carry in_science_slice: true; 347,539 rows (38.3%) carry the taxa detector flag; the flags overlap on 60,635 rows.
• 0 disambiguation pages. All rows carry the index delivery label science, which is a routing hint, not adjudicated truth.

All four delivery priorities, for context (mutually exclusive, advisory only; the dataset repo carries p0 inline, the bucket carries all four):

Priority	Articles	Share of retained tree
p0	908,270	12.8%
p1	4,665,480	66.0%
p2	1,208,553	17.1%
p3	291,274	4.1%

Keys and schemas

Grain and key per config; every key claim below is re-verified against the published payloads each time this card is built:

• refs — one row per multistream member; key: stream_id, offset; 256,564 rows, key verified unique at build time.
• articles — one row per article with parsed citation signals; key: article_l; 5,135,607 rows, key verified unique at build time.
• categories — one row per category with citation aggregates; key: category; 529,402 rows, key verified unique at build time.
• publishers — one row per normalized publisher string; key: entity_norm; 20,000 rows, key verified unique at build time.
• templates — one row per citation-template family and parse outcome; key: medium, parse_status; 37 rows, key verified unique at build time.
• clusters — one row per deduplicated citation record with its cluster; key: dedupe_key; 1,511,504 rows, key verified unique at build time.
• streams — one row per multistream; key: stream_id; 256,564 rows, key verified unique at build time.
• disambig-nodes — one row per detected disambiguation node and base title; key: disambig_title, base_title; 310,529 rows, key verified unique at build time.
• disambig-links — one row per heuristic candidate link; key: disambig_title, base_title, candidate_title, match_kind; 723,486 rows, key verified unique at build time.
• redirects — one row per redirect edge; no unique single-row key: 11,894,129 edges over 11,890,442 distinct source titles; 2,315 source titles (mostly short punctuation fragments) carry more than one target edge (3,687 surplus rows).
• articles-text-p0 — one row per retained p0-slice article record; key: stream_id, source_member, verified unique at build time; per-row provenance anchors (article_sha256, source_archive_sha256) bind every row back to the reviewed compact corpus, and row counts and shard hashes are fixed by articles-text-p0/manifest.json.

Full Arrow schema per config (field names, types, nullability):

• schema/refs.json
• schema/articles.json
• schema/categories.json
• schema/publishers.json
• schema/templates.json
• schema/clusters.json
• schema/streams.json
• schema/disambig-nodes.json
• schema/disambig-links.json
• schema/redirects.json

The p0 shard schema is documented in data/articles-text-p0/manifest.json and in each shard itself.

Load recipes

datasets (one config at a time; the p0 tier streams):

from datasets import load_dataset

refs = load_dataset("robbd/wikiparse-2606", "refs", split="train")
p0 = load_dataset("robbd/wikiparse-2606", "articles-text-p0", split="train", streaming=True)

pandas and PyArrow read the Parquet directly over the hub filesystem:

import pandas as pd

refs = pd.read_parquet("hf://datasets/robbd/wikiparse-2606/data/refs/0000.parquet")

import pyarrow.parquet as pq

refs = pq.read_table("hf://datasets/robbd/wikiparse-2606/data/refs/0000.parquet")

DuckDB queries individual p0 shards in place:

SELECT count(*) FROM
'hf://datasets/robbd/wikiparse-2606/data/articles-text-p0/shards/000000-000256/p0.parquet';

Point DuckDB at single shards or narrow shard ranges. A single * glob across all 1,003 shards resolves through the hub's tree API and reliably hits its rate limit; for full-tier scans use the streaming datasets recipe above.

The other priorities live on the mirror bucket under context/content/v1/shards/<stream range>/<priority>.parquet; see the bucket's own README for its manifest and receipt hashes.

Context delivery

The public robbd/wikiparse-2606-data bucket holds the larger, mutable delivery surface: the full four-priority article-text tree (7,073,577 rows in 1,003 disjoint shards). The dataset repo remains the smaller, versioned catalog and additionally carries the p0 tier inline as the articles-text-p0 config.

The content manifest is bound by SHA-256 (f3de8e5b6c4f0a2d26f5f17d27a915b6964c16c69707a38de24af7d6647569c6) and the publication receipt by 71bb9b65bab44d6b9237a0b6da400a653f547d420fdcae4e4172a9818f12147e. The inline articles-text-p0 copies are byte-identical to those reviewed shards; data/articles-text-p0/manifest.json carries the per-shard hashes and the tier definition.

Field bundles

bundles/ holds 44 curated Markdown zips built from the same dump by the release packer in the wikiparse repository's scripts tree: 38 single-field zips and 6 division roll-ups, covering 415,766 articles (99.47% of a 418,016-title bundle input) with their citation partitions. Each zip carries its own slice-manifest.json (selection rule, seed categories, recall/precision floors, sentinel checks) and a per-scope README.md.

Scope	Count	Payload
Fields	38	~3.4 GB
Divisions	6	~1.4 GB
Total	44	~4.78 GB

Every article in every bundle carries a resolved is_biography flag -- true, false, never a silent default. A deterministic, additive backfill against each title's own birth/death/living-people categories resolved all but 37 genuinely ambiguous multi-subject title collisions, and this bundle input carries zero unresolved identities (identity_unknown: 0 in every slice-manifest.json). A future release with a nonzero count would file those rows under an unknown-identity/ folder and say so in the bundle's own README, never fold them into articles/.

These are prepared candidate slices: the selection rule is deterministic (category-graph closure plus lead-sentence and infobox heuristics, no model calls), scored against a fixed sentinel set per field, but has not had a per-field human domain review. Read a bundle's own README before treating its membership as authoritative for a specific field.

Analysis

5 small, source-bound cards summarize the least ambiguous observations:

• category-structure
• citation-quality
• provenance-repair
• science-closure
• wave0-landscape

What this is not

• Parser signals such as suspect are operational labels, not adjudicated truth.
• Category counts overlap. Do not read their sums as distinct population totals.
• Disambiguation links are heuristic candidates; they are not an authoritative taxonomy.
• Citation records are article-level and deduplicated. They are not inline claim markers.
• The p0 delivery label science is a routing hint from an advisory index; science precision on this corpus is unmeasured, and this release does not claim it.
• 3 p0 rows exceed the dataset viewer's 2 MB per-row rendering limit (maximum article length 4,929,718 characters; 136 rows exceed 500 KB). They load normally in datasets/PyArrow/pandas/DuckDB; the viewer may not render those rows.
• This is a technical milestone, not a paper, benchmark leaderboard, or announcement.

Provenance

• Source: the processed enwiki-20260601 corpus of English Wikipedia; derived work is distributed under CC BY-SA 4.0 (see License).
• The ten metadata tables are byte-identical to reviewed local objects. provenance.json binds the reviewed source inventories, each public file, its row count, and its Arrow schema.
• articles-text-p0 shards are byte-identical copies of reviewed content shards; the tier definition, per-shard SHA-256, row counts, and the bound content and index manifests are in data/articles-text-p0/manifest.json.
• checksums.sha256 covers the complete repository tree, including the LFS-backed objects.
• Parser and release tooling: https://gitlab.com/doering-ai/apps/wikiparse

Citation

If the dataset is useful to you, cite it and attribute Wikipedia:

@misc{wikiparse2606,
  title  = {wikiparse 2606: citation, category, and article-text tables from the
            English Wikipedia 2026-06-01 dump},
  author = {Doering, Robb},
  year   = {2026},
  url    = {https://huggingface.co/datasets/robbd/wikiparse-2606},
  note   = {Derived from English Wikipedia; CC BY-SA 4.0.}
}

Text citation: Robb Doering (2026). wikiparse 2606: citation, category, and article-text tables from the English Wikipedia 2026-06-01 dump. Hugging Face dataset.

This dataset is derived from English Wikipedia, contributed by Wikimedia contributors. Redistribution and adaptation must credit the Wikimedia contributors and share alike under CC BY-SA 4.0. No Wikimedia endorsement is implied.

License

The data is derived from English Wikipedia and is distributed under CC BY-SA 4.0. See LICENSE.md for attribution and modification notes.