jhdlee/wiki-events-cpt

Wikipedia Events CPT

jhdlee/wiki-events-cpt is a public research dataset of 150 selected English Wikipedia articles with compact metadata for continual pretraining (CPT).

Split	Articles	Event window (end exclusive)
cohort_a	75	2023-01-01 to 2024-10-01
cohort_b	75	2024-10-01 to 2025-09-01

Each cohort has 25 articles per topic: natural_hazards, elections, and sports. Cohorts group events by their reviewed whole-occurrence intervals; they are not training/evaluation splits. Article revision dates do not determine cohorts.

Both cohorts use the same August 2025 text snapshot from HuggingFaceFW/finewiki, en/train, pinned to revision 8bd13e72e6a002407649b3e898535f42ceb1aeb9. Articles can contain older background and later updates, including opposite-cohort knowledge.

text preserves the exact exported CPT: original title + two newline characters + unchanged published body. Selection and this title serialization do not rewrite facts, repair source errors, translate annotations, or shorten articles.

Fields	Meaning
id, title, text	Stable article identifier, original title, and complete CPT text.
cohort, topic, subtype	Event-time cohort, topic, and event subtype.
event_start, event_end_exclusive	Reviewed whole-occurrence UTC bounds; the end is exclusive. Date-only evidence can yield conservative bounds when the original time zone is unspecified.
event_date_precision, event_time_basis	Lists of strings describing source date precision and time basis.
article_revision_date	Source article modification timestamp, separate from event dates.
narrative_tokens, cpt_tokens	Paragraph-view token count and full CPT token count, respectively.
source_url, source_revision_url	Wikipedia article and specific revision links.
page_id, revision_id	Original Wikipedia page/revision identifiers.
text_sha256, quality_flags	SHA-256 of UTF-8 CPT text and retained source-review flags.

Eligibility requires at least 2,048 narrative tokens, inclusive. The paragraph-only view excludes headings, lists, tables, code, math, and reference-like sections under the recorded policy. Full CPT retains those source elements and can be substantially longer. Selected narrative lengths have mean approximately 6,859 and minimum 2,311 tokens; full CPT totals 2,486,428 tokens, with mean 16,576 and range 2,356–135,672. Counts use the pinned authenticated tokenizer with no truncation or added special tokens. The current threshold projects preserved measurements onto 2,048 tokens without changing text or the metric.

The bundle manifest records the pinned source and tokenizer, policy and cohort windows, original receipt hashes, and package checksums. The dataset contains compact Parquet rows, this card, a license notice, and the manifest. Detailed operational review artifacts remain local.

Articles received assistant reviews; human_approval_granted remains false. Recorded source-quality holds were excluded from selection, but absence of a hold or flag does not establish factual correctness. Twenty articles retain non-Latin-annotation review flags and their original text. The original CPT-only release performed no QA generation, model inference, or training. The QA addition below used local model inference; no student training has been performed.

All 11,175 selected-article pairs received lexical similarity comparisons; this does not establish semantic independence or global corpus deduplication. A limited scan found 31 cross-cohort full-title mentions, without complete alias, paraphrase, implicit-reference, or fact-overlap coverage. Selection is complete for this reviewed working set; full-corpus review and a refreshed full-corpus inventory at the current threshold are not certified here.

Loading

Replace COMMIT_FROM_HUB with the dataset commit to reproduce.

from datasets import load_dataset
selected = load_dataset(
    "jhdlee/wiki-events-cpt", "selected", token=True,
    revision="COMMIT_FROM_HUB",
)
cohort_a_texts = selected["cohort_a"]["text"]
cohort_b_texts = selected["cohort_b"]["text"]

Text is provided under CC BY-SA 4.0, following the pinned FineWiki processed-text license. Credit Wikipedia contributors and Guilherme Penedo / FineWiki (2025); retain attribution and indicate adaptations when redistributing. The upstream notice separately records underlying Wikipedia text's GFDL availability; see Wikipedia dump licensing.

Per-row source/revision URLs provide attribution. Contributor history is available at https://en.wikipedia.org/w/index.php?curid=PAGE_ID&action=history. This release adds selection, compact metadata, and the documented title/blank-line serialization. Upstream citation: Guilherme Penedo, FineWiki, 2025, Hugging Face Datasets, pinned dataset.

Unsplit question-answer collections

The original selected CPT configuration remains the default and its Parquet files are unchanged. Two additional configuration dropdowns contain model-generated single-fact questions derived from the 150 selected full CPT articles at commit 9fcbf219592fef3dc0fb71408a8d7ec259822d5f.

Configuration	Total QAs	A	B	QAs per article
qa_40	6,000	3,000	3,000	Exactly 40
qa_all	8,504	4,313	4,191	40–60

qa_all retains every quality-passing QA remaining after within-article deduplication. It does not include rejected QAs, unresolved quality-review cases, or articles from the CPT reserves. qa_40 is a uniform random sample without replacement from each article's qa_all pool. Seed 42 is combined with the article ID using SHA-256; sampling operates on sorted QA IDs. The precise algorithm, Python version, selected IDs, hashes, and counts are recorded in qa_selection_manifest.json and qa_release_manifest.json.

Each configuration has one data storage split. There are no train, validation, or test assignments. Both configurations contain the same seven fields: qa_id, article_id, title, cohort, topic, question, and answer. Join article_id to selected.id for full CPT text, dates, token counts, source/revision URLs, and attribution. Evidence quotations, reasoning traces, and verifier judgments remain separate audit artifacts and are not model-input fields in this release.

from datasets import load_dataset

qa_40 = load_dataset("jhdlee/wiki-events-cpt", "qa_40", split="data", revision="COMMIT_FROM_HUB")
qa_all = load_dataset("jhdlee/wiki-events-cpt", "qa_all", split="data", revision="COMMIT_FROM_HUB")

Generation proposed 60 QAs per article from the full CPT, including lists and tables, using an SDFT-derived prompt adapted to request one directly stated fact about the central event per question. Generation and separate QA verification used local asynchronous SGLang with Qwen/Qwen3.8-27B-FP8 at 017b9c7af6b5689d5dd426a76e0bc077eb5ca20a, thinking enabled, native xhigh reasoning. Sampling used temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 0, and repetition penalty 1.

Checks covered response structure, source quotation matching (including audited whitespace repair), answer support, single-fact scope, self-contained wording, central-event relevance, and qualifications. Semantic deduplication was restricted to each article. Normalized question embeddings from Qwen/Qwen3-Embedding-8B at 1d8ad4ca9b3dd8059ad90a75d4983776a23d44af nominated 7,073 pairs at cosine similarity at least 0.85; 50 lower-similarity audit pairs were also reviewed. The Qwen 27B model with xhigh reasoning reviewed all 7,123 nominated pairs, yielding 13 duplicate and 7,110 distinct decisions. Thirteen QA records were removed from the eligible pool. Equivalent units or representations can describe the same fact. Different attributes or event states remain distinct. These checks do not certify exhaustive duplicate recall or factual correctness, and no cross-article semantic deduplication or human approval is claimed.

The QAs are an additional adaptation of the attributed article text, distributed with this dataset's CC BY-SA 4.0 notice. Preserve article joins for Wikipedia/FineWiki attribution. The original bundle_manifest.json authenticates the original CPT release, including its original README at the source commit; the new qa_release_manifest.json authenticates this addition.