Battam's picture
Release [run release-20261116T105650Z]
41f8a64 verified
|
Raw History Blame
10.7 kB

Back to README.md

Sources

English | 简体中文

Two layers, listed apart, whose counts are never added together and whose sizes are: task material first, corpora second. Each corpus is one whole upstream read in bulk and counted in rows; a material set is admitted for building tasks from works the search found upstream or from a group of corpus rows, its files checked and grouped into units, and it is counted in units, apart from any corpus its rows came from. Whole index: Recorded size 4.45 PB; deduplicated size 4.14 PB; size check: 99.4% match (35,599 of the 38,489 rows read compared); not compared: address unreachable 144, file gone 193, login needed 526, no size from the server 2,022, other error 5; estimated size of 278,400,014 of the 289,533,999 rows without a recorded size: 470 TB (an estimate; 95% range 323 TB to 715 TB, the sum of its sources' ranges, wider than a joint interval; from 4,388 rows read), the other 11,133,985 with too few rows read for an estimate. Task material: Recorded size 2.93 GB; deduplicated size 2.82 GB; size check: 99.5% match (11,753 of the 11,769 rows read compared); not compared: file gone 3, login needed 10, other error 3. Corpora: Recorded size 4.45 PB; deduplicated size 4.14 PB; size check: 99.4% match (23,846 of the 26,720 rows read compared); not compared: address unreachable 144, file gone 190, login needed 516, no size from the server 2,022, other error 2; estimated size of 278,400,014 of the 289,533,999 rows without a recorded size: 470 TB (an estimate; 95% range 323 TB to 715 TB, the sum of its sources' ranges, wider than a joint interval; from 4,388 rows read), the other 11,133,985 with too few rows read for an estimate. Distinct content is the sum of size_bytes over the rows minus the sizes of the rows that repeat another row's content (the same sha256, else sha1, else md5 and size, within the same hash scope; one copy of each content is kept, in the first source by registry order), counted after each release over the data files it published, per source in the catalog column distinct_bytes; archive members whose archive the same source also lists whole have other hashes, so they are not subtracted. tools/fetch.py removes duplicates over your own selection when it reads it. The deduplicated sizes here were counted over the data files of release release-2026-11-16.5. The recorded size is the sum of size_bytes over the rows that have one (catalog column bytes). The check reads a sample of rows of every data file (meta/sizes.parquet), each day the files with rows of a source not read there yet and those of the sources read longest ago, about a seventh of all the readings, so each is read again within a week: 20 of its rows with a recorded size of each source whose rows it holds, and of its rows without one its share of its source's sample, sized for their mean size within 25% at 95% confidence by the spread of the sizes read so far (at most 400 rows per source per round, 20 for a source with none read yet); its share is the rows that match their recorded size out of the rows it compared (status match, size_mismatch or hash_mismatch), and the rows it could not compare, which never reached their file, are counted beside it by status (unreachable, gone, needs_auth, no_length, error; catalog column size_check). The estimate for a source's rows without a recorded size (catalog columns unsized_rows and unsized_estimate) is their number times the mean measured size of such rows read, with a 95% BCa bootstrap range (scipy.stats.bootstrap, its default resample count); a source with fewer than 10 such rows read has none. Above one source, the estimate is the sum of its sources' estimates and its range the sum of their ranges, which is wider than a joint interval. The figures of every level are in meta/size_figures.json.

Task material

1,849 material sets with 4,403 units, in 14 collections (1 of them with no material set yet); 1 of the material sets are in no collection. On the domain pages they are listed by collection, each collection with its sizes in that domain. A material set (task-ready) is the units admitted from one source, which reads one upstream (one dataset or one data catalog), for building tasks; it is counted in units, each a group of files its upstream publishes as one thing. What a unit, a row and a collection are, and how a unit is keyed: the glossary in the data page's manual.

Corpora

139 corpora with 1,198,791,875 rows. A corpus (plural corpora; raw) is one whole upstream that a collector keeps reading; its rows point to files or records, and it is counted in rows, not units. A corpus that a collection lists as holding its kind of material (corpora in the pipeline's catalog/domains.yaml) is listed under that collection's name, still counted in rows.

Domains

Each source is listed on the page of its domain: the top-level domain of its first domain code (catalog column domain), or unclassified without one. A domain too large for one page is on numbered pages, each linking the others. The counts of the two layers stay in columns of their own, and the row of all domains gives the counts of the whole release, which the domains' rows add up to.

Domain Material sets Units Corpora Rows Pages
Software engineering 302 479 12 29,990,992 1
System administration 71 274 29 161,489,474 1
Data science 353 1,040 23 770,254,753 1
Machine learning 192 545 7 9,658,311 1
Science 178 428 15 45,649,077 1
Operations 209 466 6 39,763,158 1
Hardware 87 296 5 8,719,142 1
Security 112 368 10 4,536,079 1
Media 276 393 9 12,753,109 1
Misc 69 114 3 241,140 1
Many domains 0 0 16 109,289,735 1
Unclassified 0 0 4 6,446,905 1
All domains 1,849 4,403 139 1,198,791,875

How to read the tables

The domain pages list each source in a table of its layer, with these columns.

  • One row points at (corpora) says what a row points at: a file, an archive member, a web archive record, a byte range or an API response.
  • Rows (corpora) counts pointers in this release.
  • Units (material sets) counts the distinct unit_id values (a test case, one workbook, one package); Files counts the rows: the files of its units and the README.md and LICENSE the index writes for the source. Units and rows are never added into one number.
  • Size gives the source's figures, each saying what it is. The recorded size is the sum of the sizes of its rows that have one. The deduplicated size is that minus the sizes of its rows that repeat another row's content (the same hash; one copy of each content is kept, in the first source by registry order). The check says what share of the rows compared in meta/sizes.parquet match their recorded size, out of how many rows read, and how many rows read could not be compared, by what kept each from its file. The estimate is for the rows without a recorded size, from the measured sizes of such rows read, with its 95% range. unknown means the release does not know it; a size that was not measured is never written as 0.
  • Fetch check says whether the upstream still serves the source's files: how many of its sampled rows read in meta/sizes.parquet fetched and matched their recorded size and hash, out of how many rows read, and the day they were read (catalog column fetch_check). Each day the check reads the data files holding rows of a source not read there yet, and those of the sources read longest ago, about a seventh of all the readings, so each source is read again within a week; a release that writes a data file anew keeps the readings of the rows still in it. Bytes that are not the recorded file, such as an HTML page in place of a workbook, differ in size or sha256 and count as not fetched. A source none of whose rows fetched stays listed, with this reading beside it, and the next reading of its rows replaces it; until then it is flagged as not usable as it is (catalog column status, with the reading as its reason in status_reasons), on its card and on the data page alike. A source whose rows need a paying or signed-in reader (Login), and none of whose rows the check, which reads anonymously, could read, is at level needs auth and not flagged: that reading says nothing of whether the upstream serves them.
  • License tier groups each source's license by what it allows:
    • permissive: use, training and redistribution allowed with at most attribution (for example MIT, Apache-2.0, CC-BY-4.0, CC0-1.0).
    • share-alike: allowed, but copies and derived files keep the same license (for example GPL, CC-BY-SA); with the reason code us-pd, public domain in the United States only, so redistribution elsewhere is conditional.
    • nc-research: non-commercial or research use only.
    • no-training: the license or terms exclude training use.
    • internal-only: may be used but not redistributed.
    • mixed: rows of the source carry different tiers; the row license columns decide.
    • unknown: the source states no license, or it could not be classified; treat it as not cleared.
  • Login says whether a download needs an account, a form or a request (access.md).
  • What it is is the summary read from sample files when the source has been judged, and otherwise the first sentence of the source's own description; a material set whose description admission wrote gets this release's counts instead, its units, its files and its publishers, the same counts as the data page and its card.