URLs

74,918,894,107 deduplicated, validated URLs, sorted by SURT key and split into 2,334 range shards.

As plain text the URLs are 5.8 TiB, averaging 84 characters each. Sorted by SURT key and delta-encoded they fit in 665.6 GiB — 9.54 bytes per URL, a 8.85× reduction. That is the whole point of the ordering: SURT puts URLs from the same site next to each other, DELTA_LENGTH_BYTE_ARRAY then stores only where each row differs from the one above it, and zstd compresses what is left. The ks46/urls-sampled card shows the same URLs costing 23.89 bytes each without that adjacency.

Every number on this page is measured on the shipped files, and the artifacts that produced each figure are published alongside the data in stats/.

What this is

One URL per row, one column, no metadata. The point is coverage and order:

• Deduplicated globally. A URL appears exactly once, not once per source. 137.5 B rows from 62 datasets collapsed to 40.1 B; 119.6 B rows from 43 Common Crawl crawls added 34.8 B more.
• Sorted by surt(url), then url. SURT reverses the host (https://www.example.com/a → com,example)/a), so the corpus is ordered by registrable domain and neighbouring rows are usually the same site. That is where the compression comes from, and it is measured below.
• Range-partitioned, not hash-partitioned. Shard b holds the keys in [splits[b-1], splits[b]). A key range is a contiguous slice of the web, so a host lives in one shard — only 88 hosts in the entire corpus straddle a boundary.

For the opposite property — chunks that are uniform random samples, stable under corpus changes — see ks46/urls-sampled, the same URLs partitioned by xxh3_64 range.

The corpus in numbers

URLs	74,918,894,107
distinct hosts	490,966,709
distinct TLDs	199,713
on disk	665.6 GiB (714,741,129,463 bytes)
URL text	5.8 TiB (6,328,442,428,750 chars)
bytes per URL	9.54
mean URL length	84.5 chars
shortest / longest URL	8 / 1,063,146 chars
shards	2,334
rows per shard	min 5,596,896 · median 30,752,907 · max 92,889,462

Row counts are skewed, though far less than they were: 2 shards carry more than twice the median and none more than three times. 227 shards that ran over were split at row-group boundaries into 2,334 units — the same URLs in the same order, cut into more files, so no URL moved between shards.

The boundaries were placed on a sample of the 137.5 B raw input rows, then used to partition what those rows became: 40.1 B after global deduplication, plus 34.8 B more from Common Crawl. Deduplication does not shrink regions evenly — a URL carried by twenty sources collapses to one, a URL carried by a single source does not — so equal input rows is not equal unique rows. The split corrects the worst of that after the fact.

Bytes are far less skewed than rows, because the dense regions that end up row-heavy are exactly the ones delta encoding compresses best — a shard can hold three times the median row count in less disk than the median shard. If you are sizing downloads or scans, use bytes from shards.json rather than rows.

If you want equal-sized units of work, use ks46/urls-sampled instead: the same URLs in 2,048 hash chunks of 36,574,592 to 36,586,653 rows, a spread of 0.03%. Range partitioning and even partitioning are different goals, and each dataset does one of them.

Shape

	count	share
https://	48,288,088,009	64.5%
leading www.	31,401,491,135	41.9%
has a query string	17,698,714,684	23.6%
percent-encoded	6,793,752,283	9.1%
punycode host	155,900,167	0.21%
has a fragment	129,264,821	0.17%
explicit port	78,358,147	0.10%
non-ASCII	41,830,688	0.06%
IP-literal host	6,992,145	0.01%

What sorting buys

Share of rows whose URL has the same first n characters as the row before it:

prefix	rows	share
16 chars	62,724,910,920	83.7%
32 chars	53,181,538,042	71.0%
64 chars	14,239,372,560	19.0%

That is the delta-encoding headroom the SURT order creates, and the reason 84.5 characters compress to 9.54 bytes. The same columns in urls-sampled show what happens without it.

Largest registrable domains

domain	URLs
reddit.com	1,395,461,675
redd.it	798,711,317
wordpress.com	383,348,427
imgur.com	154,600,083
google.com	132,627,387
wikipedia.org	128,464,980
fc2.com	115,660,163
archive.org	96,740,069

Registrable domains come from the Public Suffix List. Where a registry ends and a site begins is policy, not syntax, so there is no way to derive it from the string. Taking the last two labels instead — the obvious shortcut — merges every site under a multi-label suffix into one row:

suffix	URLs	distinct sites underneath
co.uk	1,072,046,156	5,311,309
blogspot.com	981,011,663	17,358,886
com.br	648,351,694	3,215,936
com.au	453,051,282	1,947,119
co.jp	271,546,430	479,071
com.cn	242,312,545	8,814,216
ac.uk	154,167,981	54,292

Those 17.4 M Blogspot blogs would have become a single blogspot.com entry. With the list, www.bbc.co.uk resolves to bbc.co.uk, x.github.io to x.github.io, and s3.amazonaws.com to itself, since it is a suffix in its own right. stats/public_suffixes.parquet has all of them; stats/hosts_psl.parquet maps every host to its suffix and domain.

Largest TLDs

TLD	URLs		TLD	URLs
com	35,033,115,289		de	2,397,599,121
org	4,069,872,273		it	1,801,668,446
ru	3,157,550,535		co.uk	1,072,046,156
net	2,659,582,082		fr	1,062,381,327

Layout

data/part-00000.parquet … part-02333.parquet    one column: url (string)
shards.json                                     per shard: bounds, rows, bytes, first_rank
stats/                                          see below
audit.jsonl                                     per-shard verification record
removed.parquet                                 every URL dropped by cleaning: url, reason, shard

Parquet V2, zstd level 9, 1,000,000-row row groups, DELTA_LENGTH_BYTE_ARRAY.

shards.json carries first_rank per shard — the number of URLs in all earlier shards — so a global row number is first_rank + offset within the shard.

Statistics

Answering "what are the most common hosts" or "what URL sits at row 40,000,000,000" costs a read of 665.7 GiB. None of those answers are large, so they are computed once and published:

file	contents
stats/corpus.json	every corpus-wide total on this page
stats/shard_stats.parquet	31 columns of per-shard detail — length percentiles, scheme/query/fragment/port counts, first and last URL. Keyed to the pre-split 2,048-shard layout; shards.json is the current index.
stats/hosts.parquet	all 490,966,709 hosts → URL count, total and min/max/mean URL length, https and query counts
stats/registrable_domains.parquet	eTLD+1 via the Public Suffix List
stats/public_suffixes.parquet	URLs and hosts per public suffix
stats/tld.parquet	exact URL count for each of 199,713 TLDs
stats/length_hist.parquet	exact log₂ histogram of URL length
stats/path_depth.parquet	exact distribution of path depth
stats/extensions.parquet	top 5,000 trailing extensions
stats/query_param_count.parquet	exact distribution of parameters per URL
stats/query_keys.parquet	1,098,846 query parameter names
stats/query_values.parquet	top 2 M key=value pairs, value kept with its key
stats/query_value_len.parquet	value-length histogram
stats/fragments.parquet	top 500 K #fragment values
stats/rank_index.parquet	every 10,000th URL with its global rank — 7,493,418 rows, so a rank lookup is a binary search rather than a scan

Everything is exact except query_keys, query_values and query_value_len, counted on a deterministic 1-in-1,000 hash band; splitting every URL's query string would have exploded roughly 67 B rows.

Two fields in shard_stats.parquet carry a known imprecision, recorded in corpus.json: hosts and ip_literal were computed on host keys that still carried :port, which affects 0.13% of URLs. hosts.parquet is port-normalised and exact.

How it was built

Sources. 62 URL datasets plus a 43-crawl Common Crawl sweep:

101_billion_arabic_words_dataset · archive-text · c4 (en, multilingual, zh-tw, noblocklist, noclean, realnewslike) · catalog · ccrawl · clean_mc4_it · colossal-oscar-1.0 · common-crawl-sample · crawlpt_dedup · culturax · culturay · dclm-baseline-1.0 · dolma (v1.5, v1.6, v1.7) · falcon-refinedweb · finemath · fineweb · fineweb-2 · fineweb2-hq · fulg · georgian-corpus · glotcc-v1 · hackernews · hplt (v1.2, v2.0) · infimm-webmath-40b · legal-mc4 · lucie-training · madlad-400 (clean, noisy) · mc4 (es, fi, nl) · mixturevitae-fineweb-permissive-multilingual · moscar · obelics · olmoe-mix-0924 · onlysports · open-web-math · oscar (2109, 2201, 2301) · redpajama (v1, v2) · reddit (comments, submissions, wenknow) · txt360 · varta · wikipedia externallinks · zyda · zyda-2

Normalization is deliberately light — the goal is a URL corpus, not a canonical-URL corpus. Whitespace and control characters are trimmed, HTML entities decoded, a scheme added where one is clearly missing, and rows that cannot be URLs dropped. Scheme case is preserved: httpS:// is a valid URL and stays as written. 0.551% of input rows were dropped, 0.110% rewritten.

Validation. Every URL was checked against url_valid(), which answers "could this be a web URL?" from RFC 3986 and RFC 1035 in one allocation-free pass, and names the rule it broke. 5,188,960 rows were removed — 0.00693% — and every one is listed in removed.parquet with url, reason and shard.

The distribution says what the damage is. This is a link extractor run over text that was never a link:

reason	rows	what it is
bad_tail_char	3,168,516	trailing markup: http://0.0.0.0/</url>, ?token=<token>
unknown_tld	867,294	example.adac, 192.internal.ip.address, ec.europa truncated from europa.eu
bad_tld	425,339	digits injected mid-label: t.0000me, the00000federalist.00000com
single_label	252,843	no dot at all: http://0-0, https://http://'
bad_host_char	140,961	emoji and punctuation hosts: https://www.‼️‼️HOLY
label_hyphen	88,200	templates: http://COMPANYX-.$domain/repos/CentOS6/
bad_ipv6	86,683	markdown: http://[1](http://i.imgur.com/CaG7u.jpg)
too_long	36,292
empty_label	25,440	truncation: http://worldofwarshi...!/pvp/overview/
label_too_long	24,270	whole sentences percent-encoded as a hostname
bad_ipv4	15,351	http://0, http://00, http://www.0.0.0/
ipv4_leading_zero	3,202	https://207.000.000.000 — browsers read these as octal
bad_userinfo, host_too_long, empty_host, bad_port, bad_punycode	385
legacy_clean	54,184	an earlier pass: whitespace, control characters, hosts with no letter or digit

Every reason bucket was sampled across seven shards and read before the rule was turned on. unknown_tld was the only genuine judgement call at 867,294 rows; reading it showed documentation placeholders, truncations, typos and two URLs concatenated into one host, with the real-IDN risk inside it amounting to 51 rows.

The legacy_clean bucket is an earlier pass, folded into the same file: 54,184 rows with whitespace or control characters inside the URL, or hosts containing no letter or digit at all (http://"", http://*, http://.).

URLs over 4,096 characters were not removed: 1,472,208 of them are legitimate, including a 21,978-character auction query string and a 1,063,146-character museum-collection URL. Length is not a validity test.

One gap in this record, stated plainly. A conservation check compares every cleaned shard against its pre-clean original and excuses exactly the URLs listed in removed.parquet. Shard 0 comes up 23 short: 23 URLs that are absent from the shipped data and appear in no removal record. They are all unambiguously invalid — http://$host/api/v3/history, http://<pi4, http://'.$this->getPaypalHost().', https://pandavault.x.yupoo.(REMOVE)com/ — so excluding them is correct; an earlier pass simply failed to write down that it had. The shipped corpus is unaffected, but removed.parquet is 23 rows short of being a complete account of everything ever dropped. Every other shard balances.

Verification. Every shard passed all five structural checks, plus conservation:

• bounds — every key sorts inside its shard's range
• order — each shard is sorted by surt(url), url
• sequence — shard b's last key ≤ shard b+1's first
• distinct — no duplicate URLs
• unkeyed — every row has a key at all
• conservation — every URL from the pre-Common-Crawl corpus is still present, except exactly those listed in removed.jsonl, matched by URL rather than by count

distinct within a shard gives global uniqueness here, because bucket assignment is a pure function of the URL: every copy lands in the same shard.

Usage

from datasets import load_dataset

# stream — the whole corpus is 665.7 GiB
ds = load_dataset("ks46/urls", split="train", streaming=True)
print(next(iter(ds)))

# one shard, without downloading the rest
import pandas as pd
df = pd.read_parquet("hf://datasets/ks46/urls/data/part-00000.parquet")

-- the statistics, without touching the corpus
SELECT domain, urls FROM 'hf://datasets/ks46/urls/stats/registrable_domains.parquet'
ORDER BY urls DESC LIMIT 20;

Because the corpus is sorted and shards.json publishes each shard's key range, a lookup for a host reads one shard: find the shard whose range contains surt(url), then binary-search within it.

Licensing and attribution

The URLs are collected from the public datasets listed above, each under its own license, and from Common Crawl. A URL is a factual reference to a public resource; this dataset contains no page content. Redistribution here does not change any source's terms — if you rely on a particular source, check its license. Removal requests: open a discussion on this repository.

Common Crawl data is provided under the Common Crawl Terms of Use.