OwnedByDanes/Usenet-Corpus-1980-2013

Usenet Corpus 1980–2013

One of the largest curated Usenet archives in existence — 103.1 billion tokens of human discourse spanning three decades of internet history.

⚠️ Note on size indicator: The size badge above reflects the sample files hosted in this repository (65K rows). The full corpus is 103.1B tokens across 408M posts. See Dataset Summary below for complete statistics.

Sample files are free to download with no approval required. The full corpus is available for licensing — see Licensing & Access below.

---

Dataset Summary

This dataset contains deduplicated, sanitized Usenet posts from 1980 through 2013, sourced from one of the largest privately held Usenet corpora in existence. It covers thousands of newsgroups across virtually every major hierarchy (talk.*, sci.*, comp.*, rec.*, soc.*, alt.*, misc.*, news.*, humanities.*, and more), capturing the full arc of pre-social-media internet discourse — from early academic and technical discussion in the 1980s through the peak of consumer Usenet in the 1990s and 2000s.

Stat	Value
Total tokens	103,104,873,384 (103.1B, cl100k_base)
Total records	408,236,288
Total newsgroups	18,347
Date range	1980 – 2013
Format	JSONL, gzip-compressed (.jsonl.gz)
Compressed size	~141 GB
Primary language	English (96.6%)
Languages detected	100+

---

Why This Dataset

Usenet is a uniquely valuable corpus for language model training and linguistic research:

• Pre-web, human-only content. All posts predate the LLM era by decades, making this entirely human-generated text with no AI contamination.
• Longitudinal depth. Thirty-plus years of continuous discourse allows models to learn how language, culture, technical terminology, and social norms evolved over time.
• Topic breadth. From astrophysics to religious philosophy to Unix systems administration to recipe exchange — the newsgroup hierarchy provides natural topical structure across thousands of domains.
• Conversational structure. Threaded discussions, replies, and debates provide rich examples of multi-turn reasoning, argumentation, and clarification.
• Demographic diversity. Early Usenet skewed toward technical and academic users; by the 1990s it encompassed a broad cross-section of English-speaking internet users worldwide.

---

Newsgroup Hierarchies

The corpus is organized by top-level newsgroup hierarchy. Each hierarchy was processed and validated independently.

Hierarchy	Records	Newsgroups	Tokens	Avg Tok/Post	English %	Notes
alt.*	228,761,734	15,288	58.8B	257	95.8%	Largest hierarchy; highest topical diversity
rec.*	77,233,777	919	16.5B	214	99.7%	Recreation, hobbies, sports, arts
comp.*	49,990,816	1,205	10.3B	206	99.7%	Computing, software, hardware
soc.*	22,386,509	341	8.2B	367	82.8%	Social topics; highest non-English % due to soc.culture.* groups
sci.*	11,996,714	237	3.3B	276	99.4%	Science and research
misc.*	10,866,446	242	2.7B	249	99.7%	Miscellaneous topics
news.*	2,065,102	60	1.7B	828	98.8%	Usenet meta, announcements; highest avg post length
talk.*	4,308,542	47	1.3B	313	99.1%	Debate and discussion
humanities.*	626,648	8	0.2B	352	99.6%	Literature, philosophy, classics
Total	408,236,288	18,347	103.1B	253	96.6%

---

Languages

Language detection was performed on every record using Meta's fasttext LID-176 model. The corpus is predominantly English, with meaningful representation of over 100 languages. The soc.* hierarchy has the highest non-English proportion (17.2%), reflecting the international character of soc.culture.* newsgroups.

Language	Records	% of Corpus
English	394,350,819	96.60%
Polish	3,719,737	0.91%
Dutch	2,596,453	0.64%
Spanish	2,124,609	0.52%
French	1,669,450	0.41%
Italian	743,411	0.18%
German	638,250	0.16%
Russian	366,240	0.09%
Japanese	268,684	0.07%
Indonesian	218,168	0.05%
Portuguese	205,489	0.05%
Chinese	163,198	0.04%
Romanian	112,149	0.03%
Esperanto	109,382	0.03%
Croatian	97,528	0.02%
Turkish	78,543	0.02%
Afrikaans	78,199	0.02%
Serbian	57,306	0.01%
Hungarian	53,320	0.01%
Vietnamese	50,875	0.01%
80+ additional languages	~200,000	<0.01% each

---

Schema

Each record is a single Usenet post serialized as a JSON object:

{
  "text": "Post body content...",
  "group": "talk.religion.misc",
  "date": "1995-03-14",
  "subject": "Re: The nature of consciousness",
  "author": "Jane Smith",
  "id": "msg-a3f1c2e4b5d6e7f8a1b2c3d4e5f6a7b8"
}

Field	Type	Description
text	string	Full post body. Email addresses replaced with [email] token.
group	string	Newsgroup name (e.g. comp.lang.python)
date	string	Post date, ISO 8601 format where available
subject	string	Thread subject line
author	string	Display name only. Email address stripped; bare addresses replaced with [redacted].
id	string	Anonymized Message-ID. SHA-256 hashed (msg-<hex32>). Preserved for deduplication. Omitted if source ID was null or empty.

---

Data Cleaning

The corpus was processed through a documented, reproducible two-stage cleaning pipeline with validation at each stage.

Stage 1 — Pre-filtering (hierarchy-level exclusions)

Prior to record-level processing, entire newsgroup hierarchies were excluded from the corpus:

alt.binaries.* removed — This hierarchy consisted almost entirely of UUencoded and base64-encoded binary file attachments (software, images, archives). It contains negligible natural language content and was excluded in full before any further processing.

Adult content newsgroups removed — Newsgroups whose primary purpose was the distribution of adult or pornographic content were identified and excluded at the hierarchy level. This was done as a deliberate curation decision to produce a corpus suitable for general-purpose AI training without adult content filtering requirements downstream.

These exclusions explain the low binary drop rates observed in the record-level pipeline — the vast majority of binary content had already been removed at this stage.

Stage 2 — Record-level cleaning

Deduplication — Posts with identical Message-IDs were removed. Cross-file deduplication was performed prior to this release.

Binary removal — Posts containing UUencoded attachments (begin NNN filename blocks), UUencoded data lines, or base64 binary blocks (4+ consecutive 60-char encoded lines) were dropped entirely. Binary payload detection used MIME header inspection.

PII redaction — author field — All angle-bracket groups (<...>) are stripped unconditionally from the From header — these always contain an address or Message-ID in Usenet format. The remaining display name is kept; records with no display name become [redacted]. A final regex pass ensures no email-format string survives in any edge-case format.

PII redaction — text field — All email addresses in post bodies (including obfuscated variants such as user at domain dot tld) were replaced with the [email] token.

Message-ID anonymization — Message-IDs (which are email-format strings by the Usenet protocol spec) were SHA-256 hashed to a 32-character hex prefix (msg-<hex32>). Deduplication utility is preserved since the hash is deterministic.

Sensitive content removal — Posts matching spam and bulk-PII patterns (SSNs, credit card numbers, passport numbers, bulk solicitation language) were dropped entirely.

Quoted text — Usenet reply quoting (> prefix lines) was preserved in this release as it provides valuable context for conversational structure. This can be filtered by consumers if desired.

Cleaning outcomes (Stage 2 — record-level, full corpus)

Hierarchy	Records In	Records Out	Dropped	Drop %
alt.*	229,089,954	228,761,734	328,220	0.143%
rec.*	77,272,701	77,233,777	38,924	0.050%
comp.*	50,019,440	49,990,816	28,624	0.057%
soc.*	22,398,864	22,386,509	12,355	0.055%
sci.*	12,013,350	11,996,714	16,636	0.138%
misc.*	10,880,231	10,866,446	13,785	0.127%
news.*	2,074,124	2,065,102	9,022	0.435%
talk.*	4,311,576	4,308,542	3,034	0.070%
humanities.*	627,038	626,648	390	0.062%
Total	408,687,278	408,236,288	450,990	0.110%

Drop reasons across full corpus: UUencode lines (126,441) · Base64 binary (75,164) · UUencode blocks (16,315) · Sensitive content (232,626).

Validation

After cleaning, all files were validated with a full-corpus scan (--sample 0) using a purpose-built validator checking for email addresses, obfuscated emails, binary content, sensitive content, and duplicate IDs.

Result: 15,139/15,288 files (99%) clean with zero email or PII violations. Remaining flagged files contain only preserved Usenet quote lines (> prefix) and a negligible number of duplicate IDs, both considered acceptable for training use.

---

Intended Use

This dataset is intended for:

• Pre-training and continued pre-training of large language models, under a licensing agreement
• Domain adaptation for models targeting conversational, technical, or historical internet text, under a licensing agreement
• Linguistic research on the evolution of written English over the internet era
• Retrieval and search system development using historically grounded text
• Evaluation and experimentation using the freely available sample files

Out-of-scope use

• This dataset should not be used to re-identify individuals or reconstruct contact information
• Automated profiling or targeting of individuals mentioned in posts is prohibited under the license terms
• Use in systems designed to generate spam, harassment, or disinformation is prohibited

---

Licensing & Access

This dataset is released under a custom license. The terms differ between the sample files and the full corpus:

Sample files (available in this repository):

• Free to download — no approval required
• Permitted for evaluation, non-commercial research, and personal experimentation
• Fine-tuning or training on the sample files is permitted for non-commercial, experimental use

Full corpus (103.1B tokens, 408M posts):

• Available under a separate written licensing agreement
• Commercial AI training use requires a licensing agreement
• Research institution licensing available for academic use
• Redistribution or sublicensing of the full corpus is not permitted without written agreement

To discuss licensing or data partnerships: Contact OwnedByDanes via HuggingFace, open a Discussion on this dataset page, or email usenetoverlord@gmail.com.

---

Provenance & Derivative Work

The underlying posts are public Usenet communications originally published in publicly accessible newsgroups. The compiled, processed corpus represents a derivative work produced through substantial original effort, including format conversion, deduplication, sanitization, binary removal, PII redaction, and curatorial decisions about scope and inclusion. These transformations constitute original work protected separately from the underlying source material, consistent with standard database and compilation rights frameworks.

---

Citation

If you use this dataset in published research, please cite it as:

@dataset{usenet_corpus_1980_2013,
  author    = {OwnedByDanes},
  title     = {Usenet Corpus 1980–2013},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/OwnedByDanes/Usenet-Corpus-1980-2013}
}

---

Dataset Card Authors

Dataset curated and cleaned by OwnedByDanes.