dataset_info:
features:
- name: text
dtype: string
splits:
- name: test
num_bytes: 15728640
num_examples: 540
download_size: 15728640
dataset_size: 15728640
license: cc0-1.0
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- history
- world-war
- 20th-century
- wikipedia
- rag
Chronos Wiki Corpus 📜
540 Wikipedia articles walked into a FAISS index. The retriever said "top 4." The cross-encoder said "just one, actually." The language model said "let me tell you about a Finnish shipwreck." We fixed that last part.
This is the dataset that powers Chronos-3B — a curated, cleaned collection of Wikipedia articles covering the most consequential events, figures, and inventions of the 20th century. Built for RAG. Useful for a lot more than that.
540 documents. 15 MB. The entire century, more or less.
What's inside
| Category | Examples |
|---|---|
| World War I | Causes of WWI, Western Front, Treaty of Versailles, Trench warfare, Christmas Truce |
| World War II | European Theatre, Pacific War, D-Day, Pearl Harbor, Atomic bombings, Holocaust |
| Cold War | Iron Curtain, Cuban Missile Crisis, Berlin Wall, Space Race, Nuclear disarmament |
| Inventions & Technology | Radar, Penicillin, Transistor, Computer, Internet, Jet engine, Satellite |
| Political & Social Movements | Decolonisation, Civil rights, Women's suffrage, Marshall Plan, United Nations |
| Key Figures | Churchill, Roosevelt, Stalin, Hitler, Mussolini, Truman, Eisenhower, Rommel |
The corpus focuses strictly on 1900–1999, with a heavy emphasis on the two world wars and the Cold War. No sports. No celebrity biographies. No articles about specific types of zoo keys. Just history.
How it was built
The pipeline had four stages:
1. Category-based scraping
The Wikipedia API was queried across 26 carefully chosen categories — things like World_War_I, Cold_War, 20th-century_inventions. This returned over 1,000 candidate article titles.
2. Smart filtering List articles, timelines, bibliographies, disambiguation pages, and anything about films, novels, music, or sports were automatically removed. The goal was historical substance, not breadth for its own sake.
3. Raw text extraction
For each valid page, raw wikitext was fetched via ?action=raw and converted to plain English — templates stripped, HTML removed, references gone, markup cleaned. What remains is the narrative text a reader would actually read.
4. Quality control Articles under 400 characters were discarded. Stubs and disambiguation pages produce terrible retrieval chunks. The 540 that made it through are actual, substantive articles.
Usage
For RAG pipelines
from huggingface_hub import snapshot_download
folder = snapshot_download(
"QuantaSparkLabs/chronos-wiki-corpus",
repo_type="dataset"
)
# All .txt files are in the downloaded folder
# Chunk them, embed them, index them — you're done
For fine-tuning Load the text column directly and use it as a domain-specific pretraining corpus to push a model's 20th-century historical knowledge.
For text analysis 540 structured, clean historical articles make a solid corpus for topic modelling, entity extraction, stylometric analysis, or anything else that benefits from consistent, well-written source text.
Dataset stats
| Property | Value |
|---|---|
| Total articles | 540 |
| Total size | ~15 MB |
| Language | English |
| Source | Wikipedia (public domain / CC-BY-SA 3.0) |
| Time period covered | 1900 – 1999 |
| Format | Plain text, one article per file |
| Min article length | 400 characters |
License
Source text is from Wikipedia, available under the Creative Commons Attribution-ShareAlike 3.0 Unported License. This curated collection is released under the same license. Use it, build on it, cite it.
Acknowledgments
This dataset exists because thousands of Wikipedia editors have spent years documenting history for free, for everyone. The scraping, filtering, and cleaning took days. Their work took decades. The credit belongs to them.
Built as part of the Chronos project. If you use this corpus in your own work, open a discussion on the dataset page — we'd genuinely like to know what you made with it.