slashyear — the dated historical record, with the source revision on every row

Every dated historical event we could extract from English Wikipedia's year, decade and century articles, spanning from roughly 3000 BCE to the present.

The point of this dataset is the last column. Every row carries source_revid, the numeric id of the exact Wikipedia revision the sentence was quoted from. A Wikipedia article is a moving target — a quotation you take today may not be there in six months, and a reader who follows your link cannot see what you actually read. A revision id names one immutable version that will read the same way in ten years. So a model that answers from this data can hand the user a link that stays true, which is not something you get by quoting an encyclopedia article.

Live API, full-text search and MCP server: https://slashyear.com/data

Configs

config	one row is
events	one dated event, quoted verbatim, with its source revision
years	one year, with its lead summary
subjects	one subject (person, place, institution), most with a Wikidata QID
subject_events	the join: one (subject, event) pair, so "every dated line mentioning Rome" is a filter

from datasets import load_dataset

events = load_dataset("<repo-id>", "events", split="train")
events.filter(lambda r: r["year"] == 1969)

rome = load_dataset("<repo-id>", "subject_events", split="train")
rome.filter(lambda r: r["slug"] == "rome")

Fields

events: year (astronomical numbering — -43 is 44 BCE), year_label, date, topic, text, source_title, source_revid, source_section, source_url, page.

subjects: slug, label, qid, description, entries, years, first_year, last_year, wikipedia, wikidata, page. The rows carrying a qid are joinable against any other database keyed on Wikidata, with no name matching.

subject_events: slug, label, qid, year, year_label, date, topic, text, source_revid, source_url.

How it was built

Deterministic extraction from the wikitext of English Wikipedia's year, decade and century articles at a pinned revision, then mechanical cleaning. No language model touches the wording. A model does exactly one thing: choose which of twelve topic labels a sentence files under. That decision cannot make a row say anything untrue, because it never alters the row.

The subject associations are not inferred either — a subject is attached to an event because a Wikipedia editor wrote a [[wikilink]] to it inside that dated line.

Limitations, stated plainly

This dataset inherits Wikipedia's errors. If a date is wrong there it is wrong here. What it does not inherit is Wikipedia's instability, and it adds no errors of its own, because no step in the pipeline is permitted to generate text. Coverage follows what Wikipedia's year articles happen to cover, which is uneven across periods and heavily weighted toward Europe and the last three centuries. Entries before roughly 1000 BCE are sparse and often approximate. English only.

Licence and attribution

CC BY-SA 4.0, inherited from Wikipedia. Free for any use including commercial use and including training. Every row carries the article title and revision id needed to attribute automatically:

{source_title}, English Wikipedia, revision {source_revid}. {source_url}