directorybear/open-library-10k

Dir Bear Open Library — 10,000 distilled web documents

Ten thousand complete, cleaned, English web documents — every one at least 250 words of prose, boilerplate stripped, exact-deduplicated, token-counted and scored by the quality of the site it came from. This is the free, open slice of the Dir Bear corpus: the same records, the same schema and the same pipeline as the datasets we sell, at a size you can read through in an afternoon.

• Browse it online: https://dirbear.com/library/ (every document has a reader page and a .md file)
• One file: all.jsonl.gz in this repo, or curl -O https://dirbear.com/library/all.jsonl.gz
• Licence: CC BY 4.0 — commercial use is fine with attribution

from datasets import load_dataset
lib = load_dataset("directorybear/open-library-10k", split="train")
lib.filter(lambda r: r["bear_rank"] >= 75 and r["shelf"] == "science-engineering")

What "distilled" means

Public web crawl text goes through six gates and only what survives is kept — nothing is truncated, summarised or rewritten:

gate	what it does
1 · Domain	every source domain is looked up in the Domain Index and scored 0–100 (Bear Rank); parked domains and link farms are cut
2 · Safety	domains on malware / phishing / adult / gambling blocklists are dropped, plus a conservative unsafe-term check on the text
3 · Extract	cookie banners, nav crumbs, share bars, "you might also like", copyright footers and bylines are removed line by line
4 · Substance	English only, ≥ 250 words, prose-like (list ratio, words per line, stop-word ratio) — this is where most pages die
5 · Dedup	exact-duplicate text is removed
6 · Label	tiktoken (cl100k) token count, Flesch–Kincaid reading level, cleanliness score, title, stable id and a topic shelf with a confidence

On the crawl this library was drawn from, roughly half the pages and a third of the tokens were removed. Documents were selected round-robin across shelves, highest shelf confidence first, so every shelf is represented.

Schema

One row per document; identical to the licensed Dir Bear datasets.

column	type	meaning
id	string	blake2b of the source URL — stable across builds
url, domain	string	where the text came from
dump, crawled	string	crawl dump id and crawl date
bear_rank	int	0–100 quality score of the source domain
tier	string	grizzly ≥ 90 · kodiak 75–89 · polar 50–74 · cub < 50
shelf	string	science-engineering · software-docs · health · finance · education · home-craft · history-culture · civic-law · general
shelf_confidence	float	0–1; general means the document could not be placed
title	string	first short line of the page, else derived from the URL
words, tokens	int	word count and exact cl100k token count
reading_level	float	Flesch–Kincaid grade
cleanliness	float	share of prose lines after boilerplate removal, 0–100
family_safe	bool	source domain's safety sub-score ≥ 90
build	string	Dir Bear build id
text	string	the complete document, paragraphs separated by blank lines

Shelves in this release

shelf	documents	words	tokens
science-engineering	1,250	909,446	1,180,329
software-docs	1,250	824,318	1,045,206
health	1,250	579,401	733,831
finance	1,250	591,609	747,557
education	1,250	575,252	718,656
home-craft	1,250	585,868	749,642
history-culture	1,250	544,079	718,976
civic-law	1,250	570,102	716,615
total	10,000	5,180,075	6,610,812

The rest of the corpus

The Open Library is 10,000 documents. The full Dir Bear corpus is the same pipeline run over the newest public crawls — millions of documents, sold by shelf (one-time licence) or as the Whole Library (twelve months, every new build included), delivered through the dirbear download client as zstd-compressed JSONL with local Parquet conversion. Details and pricing: https://dirbear.com/datasets/. Questions: hello@dirbear.com.

Licence and attribution

CC BY 4.0. Please cite as "Dir Bear Open Library, https://dirbear.com" and keep NOTICE.txt with the data. Source pages keep their own copyright; each record carries its URL, crawl dump and date. Web text derives from public Common Crawl data via the FineWeb dataset (HuggingFaceFW/fineweb, ODC-By 1.0).