You need to agree to share your Hugging Face information to access this dataset
Requests are reviewed manually. Please tell us who you are and what you plan to do with the data; requests without this information may be rejected.
This is a proprietary dataset. Access is granted for evaluation only and conveys no right to use, copy, or train on the data. Any use requires a separate written license from OwnedByDanes LLC (commercial, or non-commercial academic). See the LICENSE file. To request a license, contact todd@ownedbydanes.com.
Log in or Sign Up to review the conditions and access this dataset content.
Usenet Corpus 1980–2013
A large, cleaned corpus of Usenet posts spanning 1980–2013 — long-form, pre-web conversational text across the full newsgroup hierarchy. A thread- reconstructed companion (Usenet Corpus 1980–2013 — Threaded) adds conversation structure over identical content.
▶ Free preview — no gating. Browse a showcase sample of complete conversations at Usenet-Corpus-1980-2013-Full-Samples — no access request needed.
▶ License the full corpus. This dataset is available for commercial use — training, fine-tuning, evaluation, and retrieval. To discuss access and terms, contact todd@ownedbydanes.com.
Figures below are the corpus as distributed — after withholding 12 newsgroups for a documented false-positive scan match (see Distribution & omissions). The fully processed corpus, before that omission, totaled 408,033,020 posts (103.093B tokens) across 18,347 newsgroups; the Contents by hierarchy table further down reflects that pre-omission processing run and has not been recomputed row-by-row for the 12 omitted groups (see the note under that table).
- Posts: 405,794,692
- Tokens: 102,517,230,418 (102.52B,
cl100k_base) - Newsgroups: 18,335
- Distinct posts after de-duplication: 343,700,814 (15.77% of posts are cross-post copies) — measured on the pre-omission corpus; not yet independently recomputed for the as-distributed post count (expected to shift only marginally, since the omission affects 0.55% of posts, but not yet verified — see De-duplication).
- Format: gzipped JSON Lines, one post per line, one file per newsgroup
- Primary language: English (~95–99% per hierarchy; smaller Polish, Dutch, and other-language segments, chiefly within ALT)
Post counts are raw (a message cross-posted to k groups appears in k files). Cross-posting is measured; see De-duplication for the distinct figure.
Contents by hierarchy
These figures describe the fully processed corpus prior to the 12-newsgroup omission below (18,347 newsgroups, 408,033,020 posts). They have not been recomputed per row for the omission, which removes 5 ALT groups, 2 COMP groups, 4 REC groups, and 1 NEWS group (2,238,328 posts / 0.576B tokens corpus-wide, about 0.55% of the total). The as-distributed totals — 18,335 newsgroups, 405,794,692 posts, 102.517B tokens — are the figures to cite for what's actually licensed; see Distribution & omissions.
ALT is split into two equal-by-name halves for distribution
(usenet-alt-part1, usenet-alt-part2), divided alphabetically at
[fill: last group of part1, e.g. "alt.k…"]. Percentages are token share.
| Hierarchy | Newsgroups | Posts | Tokens | Tokens (B) | % of Corpus | Avg Tok/Post | Date Range | Compressed | Uncompressed | Ratio |
|---|---|---|---|---|---|---|---|---|---|---|
| ALT-PART2 | 7,644 | 136,065,542 | 34,954,698,086 | 34.955 | 33.91% | 256.9 | 1984–2013 | 49.18 GB | 152.93 GB | 3.11x |
| ALT-PART1 | 7,644 | 92,497,667 | 23,844,798,822 | 23.845 | 23.13% | 257.8 | 1981–2013 | 32.62 GB | 104.21 GB | 3.19x |
| REC | 919 | 77,237,482 | 16,501,957,075 | 16.502 | 16.01% | 213.7 | 1984–2013 | 26.38 GB | 78.07 GB | 2.96x |
| COMP | 1,205 | 49,973,522 | 10,274,541,738 | 10.275 | 9.97% | 205.6 | 1980–2013 | 17.10 GB | 50.55 GB | 2.96x |
| SOC | 341 | 22,387,270 | 8,220,160,115 | 8.220 | 7.97% | 367.2 | 1980–2013 | 11.79 GB | 34.86 GB | 2.96x |
| SCI | 237 | 11,996,904 | 3,305,179,761 | 3.305 | 3.21% | 275.5 | 1986–2013 | 5.46 GB | 16.07 GB | 2.94x |
| MISC | 242 | 10,867,041 | 2,706,808,695 | 2.707 | 2.63% | 249.1 | 1986–2013 | 4.41 GB | 13.30 GB | 3.02x |
| NEWS | 60 | 2,072,402 | 1,716,989,042 | 1.717 | 1.67% | 828.5 | 1980–2013 | 2.04 GB | 6.50 GB | 3.19x |
| TALK | 47 | 4,308,542 | 1,347,092,214 | 1.347 | 1.31% | 312.7 | 1986–2013 | 1.88 GB | 5.74 GB | 3.06x |
| HUMANITIES | 8 | 626,648 | 220,740,585 | 0.221 | 0.21% | 352.3 | 1992–2013 | 374.27 MB | 1.05 GB | 2.79x |
| Processed total (pre-omission) | 18,347 | 408,033,020 | 103,092,966,133 | 103.093 | 100.00% | 252.7 | 1980–2013 | 151.24 GB | 463.27 GB | 3.06x |
| Distributed total (as licensed) | 18,335 | 405,794,692 | 102,517,230,418 | 102.517 | — | — | 1980–2013 | — | — | — |
Languages
Content is predominantly English (~96.6% of posts). Language was detected automatically per post with fastText LID-176; figures are record counts on the raw (pre-de-duplication, pre-omission) corpus, not token shares. Automated detection on short, quoted, or code-switched Usenet text is error-prone, so treat these as approximate, especially the lower-count languages, which include some misclassification.
Languages with at least 100,000 posts:
| Language | Code | Posts |
|---|---|---|
| English | en | 394,350,819 |
| Polish | pl | 3,719,737 |
| Dutch | nl | 2,596,453 |
| Spanish | es | 2,124,609 |
| French | fr | 1,669,450 |
| Italian | it | 743,411 |
| German | de | 638,250 |
| Russian | ru | 366,240 |
| Japanese | ja | 268,684 |
| Indonesian | id | 218,168 |
| Portuguese | pt | 205,489 |
| Chinese | zh | 163,198 |
| Romanian | ro | 112,149 |
| Esperanto | eo | 109,382 |
Roughly 40 further languages appear in smaller volumes (each under 100,000 posts) and are not itemized here; those low-count labels are the least reliable.
Record schema
Each line is a JSON object:
| Field | Type | Description |
|---|---|---|
id |
string | Stable per-post identifier (msg- + SHA-256 hash of the original Message-ID). |
group |
string | Newsgroup the post was archived under. |
date |
string | Original post date (RFC 2822). A small number of records carry malformed dates. |
author |
string | Author display name as posted, with email addresses removed. Where the From header contained only an email address (about 8% of posts), the value is [redacted], so distinct-author counts understate participation. See Author field below. |
subject |
string | Subject line. |
text |
string | Post body. |
Cleaning, safety, and provenance
Binary attachments and alt.binaries.* groups were removed; adult-content groups
were excluded; author email addresses were removed; email addresses in post bodies
were replaced with an [email] token; and Message-IDs were replaced with salted
hashes. Content is otherwise preserved as posted, including period
spelling, formatting, and views.
Malware pass. The corpus was scanned with ClamAV in two stages: a per-post signature scan, then a whole-file scan with minimal-span localization for signatures spanning multiple posts (e.g. multi-part encoded binaries). Any content matching a ClamAV signature was removed. The pass is deliberately conservative: matches on quoted phishing emails or exploit code reproduced in discussion of a threat were removed along with genuine payloads, erring toward a corpus that passes the same scanner infrastructure downstream users run. All removals are logged. Removed content totals well under 0.1% of posts and is concentrated in security-, virus-, and warez-oriented groups. The distributed corpus passes a whole-file ClamAV scan with no detections.
Distribution & omissions. Twelve newsgroups are withheld from the published repository. Their historical content — security, hacking, and net-abuse discussion, plus a JavaScript programming group — matches an over-broad antivirus signature that produces a false positive on the archived text. The identical files scan clean under current ClamAV definitions (they matched only an older signature database), and were withheld so the distributed dataset passes automated security scanning without exception. The omitted groups (2,238,328 posts / 0.576B tokens corpus-wide, about 0.55% of the processed total): alt.games.tombraider, alt.games.mame, alt.hacker, alt.support.depression.medication, alt.quotations, comp.lang.javascript, comp.security.misc, rec.skydiving, rec.radio.amateur.equipment, rec.arts.comics.dc.universe, rec.music.rock-pop-r+b.1970s, news.admin.net-abuse.sightings. The headline stats at the top of this card are the as-distributed figures after this omission; the Contents by hierarchy table reflects the corpus as processed before it.
De-duplication. Cross-posting was pervasive on Usenet: a message posted to
several newsgroups is stored once per group and shares one identifier across them.
Of 408,033,020 processed posts, 343,700,814 are distinct — 15.77% are
cross-post copies. The copies are retained rather than removed, because the set
of groups a message was posted to is itself signal; the shared id lets a user
collapse them in a single pass if desired. De-duplication must be corpus-wide,
not per-file or per-hierarchy: 13.08M identifiers appear in more than one
hierarchy. Duplication is uneven by hierarchy (SOC ~18%, ALT ~15%, HUMANITIES
~1%), tracking where cross-posting was culturally heaviest. Token totals above
are likewise raw and include cross-post copies. Note: this distinct-post
count was computed before the 12-newsgroup omission above; it has not yet been
independently recomputed against the 405,794,692-post distributed corpus.
Author field. The cleaner keeps the display name from each post's From
header and removes everything else: angle-bracket groups and any email address,
including obfuscated forms. If nothing is left, the field is set to the literal
string [redacted]. In practice that means Name <addr> and addr (Name) both
become Name, while a header that held only an address becomes [redacted].
Across the posts as processed, 32,440,895 of 408,033,020 (7.95%) carry [redacted]
as their author. The share varies by year, from about 2% (2003) to 12% (1998
and 2007) among years with more than 100,000 posts, and is much higher (about
31%) among the ~2.5% of records whose date is missing or unparseable. Consequences
for users: distinct-author counts, posts-per-author statistics and participants-per-thread
figures understate real participation, and [redacted] should be excluded from any
author-level analysis rather than treated as one person. This share was measured
on the processed corpus before the 12-newsgroup omission described above; the
withheld groups hold 0.55% of posts, so the share in the distributed corpus differs by less than
one percentage point.
Missing identifiers. About 2.36M posts (~0.6%) have no recoverable Message-ID (missing or unparseable at archival). These are excluded from de-duplication.
Intended use and licensing
This is a proprietary dataset owned by OwnedByDanes LLC. It is made available
for evaluation of a potential commercial license only. Downloading or accessing
the files conveys no right to use, copy, redistribute, or train on the data;
any such use requires a separate written commercial license from OwnedByDanes LLC.
Gated-access approval is not a license. See the LICENSE file for the full notice.
To request a commercial license (including training/evaluation use for language
models), contact todd@ownedbydanes.com.
Considerations
Usenet is unmoderated historical text. It contains offensive, biased, and factually wrong content, and reflects the demographics and norms of its era. It is a snapshot of how people wrote and argued online from 1980 to 2013, and should be treated as a primary historical source, not as vetted or representative of any group's views.
Citation
@dataset{usenet_corpus_1980_2013,
title = {Usenet Corpus 1980--2013},
author = {OwnedByDanes},
year = {2026},
note = {Cleaned; commercially licensed},
url = {https://huggingface.co/datasets/OwnedByDanes/Usenet-Corpus-1980-2013-Full}
}
- Downloads last month
- 62