alvations/annotated-wiki-2016

Annotated Wikipedia 2016

English Wikipedia (~2016 snapshot) processed into JSON, with inline hyperlinks preserved as gold-aligned entity-link annotations. Each article carries its full plain text plus a list of (surface_form, target_uri, character_offset) tuples — one per wikilink in the source.

This release just re-shards the original ~5 GB stored zip (extracted/AA/wiki00 … extracted/NN/wiki47, 35,148 JSONL chunk files) into parquet with a unified pyarrow schema. No filtering, no text normalisation, no entity-set restriction. Offsets are byte-faithful character offsets into text — verified to round-trip exactly.

Mirror of the raw zip: alvations/stash · wikipedia-json/.

Usage

from datasets import load_dataset

ds = load_dataset("alvations/annotated-wiki-2016", split="train", streaming=True)
ex = next(iter(ds))
print(ex["title"], "—", len(ex["annotations"]), "links")
# Anarchism — 639 links

# offsets line up with text spans
for a in ex["annotations"][:3]:
    s = ex["text"][a["offset"]:a["offset"] + len(a["surface_form"])]
    assert s == a["surface_form"]
    print(a["surface_form"], "->", a["uri"])

Streaming is recommended — the full train split is multi-GB and 1M+ rows.

Schema

Column	Type	Description
title	string	Article title, URL-decoded from the wiki URL path.
url	string	Original Wikipedia URL, e.g. http://en.wikipedia.org/wiki/Anarchism.
wiki_id	int64	Wikipedia page ID (the source dump stored it as a 1-element list — flattened here).
text	string	Full article plain text (no markup, no infoboxes).
annotations	list<struct<surface_form: string, uri: string, offset: int32>>	One entry per outgoing wikilink in the article.
language	string	Always "en" in this release.

Annotation semantics

Each annotation marks one outgoing hyperlink in the source wikitext:

• surface_form — the literal text span as it appears in the article (the link's anchor text).
• uri — the linked target page, in Wikipedia URL form with underscores, e.g. Political_philosophy. Use as-is for joining against page-id tables; no namespace prefix.
• offset — zero-based character offset into text where surface_form begins. text[offset : offset + len(surface_form)] == surface_form is guaranteed.

Anchors are not deduplicated, redirected, or filtered — multiple links to the same target appear separately, and dangling/redirect targets are kept verbatim.

Sharding

Parquet shards are grouped by the source zip's two-letter subdirectory (extracted/<XX>/):

data/AA.parquet
data/AB.parquet
...
data/NN.parquet      (352 shards total)

Each shard is ~25 MB compressed (snappy), holding roughly 1,800–4,500 articles depending on average article length in that bucket.

Source

Re-packaged from the upstream zip wikipedia-json.zip (5.0 GB stored / 18 GB uncompressed), which was itself the output of processing an English Wikipedia dump with a wikiextractor-style pipeline that retained wikilinks as offset annotations. The dataset is named 2016 after the apparent dump year of the upstream archive.

License

Wikipedia text is distributed under CC BY-SA 3.0 (with parts also under the GFDL). Attribution: contributors to the English Wikipedia. Reuse must preserve attribution and ShareAlike.

Provenance of this release

• Source: wikipedia-json.zip — mirrored at alvations/stash · wikipedia-json/wikipedia-json.zip.
• Conversion: streamed via zipfile.ZipFile.open() without on-disk extraction; one parquet shard per source subdir; nested annotations stored as list<struct> for native pyarrow / Polars / DuckDB consumption.
• No row filtering, deduplication, redirect resolution, or text cleanup applied.