This is the complete 103.07B-token threaded corpus (≈151 GB, 18,344 newsgroups, 188.4 million reconstructed conversation threads). Access is granted under a written licensing agreement. A free, ungated sample is available at OwnedByDanes/Usenet-Corpus-1980-2013 — start there if you are evaluating.

Access approval grants file access only. It is not a license. All use — including model training, evaluation, and storage — requires a separate written agreement, as set out in the LICENSE file in this repository. Redistribution of the corpus, or of derivatives that substantially reconstitute it, is not permitted. Requests from institutional email addresses (.edu, .ac.uk, or equivalent) are prioritized. Personal email addresses (Gmail, Hotmail, etc.) will not be approved for full corpus access.

Usenet Corpus 1980–2013 — Threaded Full Corpus

The complete corpus with reconstructed conversation threads: 103.07 billion tokens, 408.08 million posts, 188.44 million threads, 18,344 newsgroups, 1980–2013.

This is the gated full release of the threaded variant of the dataset sampled at OwnedByDanes/Usenet-Corpus-1980-2013. The unthreaded sibling repository is OwnedByDanes/Usenet-Corpus-1980-2013-Full.

Evaluating? Don't request access here — use the sample. The sample files are free, ungated, and use the identical cleaning pipeline. This repository exists to deliver the corpus to licensees, not to be browsed.

Access approval is not a license. Approval grants the ability to download the files. All use — including training, fine-tuning, evaluation, and storage — requires a separate written agreement. See LICENSE.

Key Statistics

Stat	Value
Total posts (records)	408,079,574
Unique Message-IDs	343,792,481
Crosspost duplicates	61,927,304 (15.3% of records)
Total threads	188,439,945
Total tokens	103.07 billion (`cl100k_base`)
Newsgroups	18,344
Files	18,344 `.threaded.jsonl.gz` — one per newsgroup
Date range	1980–2013
Post retention vs unthreaded	99.96%
Token retention vs unthreaded	99.97%
Primary language	English (~96.6%)
Compressed size	~151 GB

Crossposts are retained. The same message may appear in more than one newsgroup file when it was originally posted to multiple groups. Headline post and token counts are record counts, not unique-message counts. See Crossposts.

Hierarchies (Threaded)

Hierarchy	Threads	Posts	Tokens	Tokens (B)	% of Corpus	Avg Tok/Post
`alt.*`	102,875,032	228,612,400	58,786,816,517	58.787	57.05%	257.1
`rec.*`	36,024,363	77,233,777	16,501,458,595	16.501	16.00%	213.7
`comp.*`	27,466,549	49,983,436	10,276,064,353	10.276	9.97%	205.6
`soc.*`	8,632,834	22,386,509	8,219,738,347	8.220	7.97%	367.2
`sci.*`	4,915,643	11,996,714	3,305,177,378	3.305	3.21%	275.5
`misc.*`	5,872,815	10,866,446	2,706,659,975	2.707	2.63%	249.1
`news.*`	1,151,535	2,065,102	1,709,091,811	1.709	1.66%	827.6
`talk.*`	1,309,812	4,308,542	1,347,092,214	1.347	1.31%	312.7
`humanities.*`	191,362	626,648	220,740,585	0.221	0.21%	352.3
Total	188,439,945	408,079,574	103,072,839,775	103.07	100%	252.6

Notes: alt.* is the highest volume and diversity. comp.* and sci.* offer the greatest technical depth. soc.* is the most multilingual. humanities.* is small (8 groups) but completes the Big 8.

What "Threaded" Means

Every post carries three additional fields that reconstruct the original Usenet conversation structure:

Field	Type	Description
`thread_id`	string or null	Deterministic conversation identifier (`thread-<sha256 of root Message-ID, first 32 hex chars>`)
`thread_position`	int or null	0-based order of the post inside its thread (BFS from the root)
`thread_depth`	int or null	Distance from the root post (`0` = root, `1` = direct reply, …)

Average thread length is approximately 2.17 posts.

Reconstruction Method

Source of truth. Original zipped MBOX files retain Message-ID and References headers.
Parent detection. The last Message-ID in References is treated as the direct parent (standard Usenet convention).
Root finding. Messages with no resolvable parent in the surviving data become thread roots.
Thread ID. thread- plus the first 32 hex characters of SHA-256 of the root’s original Message-ID.
Ordering. Breadth-first traversal from the root produces thread_position and thread_depth.
Join to cleaned posts. Original Message-IDs are hashed with the same scheme used in the cleaned corpus (msg-<hex32>), then joined onto the already-cleaned JSONL files.

Threading uses header metadata only. Raw post bodies from the MBOX files are never written into the output.

Posts that cannot be linked into a thread (missing or malformed headers) still appear, with thread_id, thread_position, and thread_depth set to null.

Cleaning artifacts in thread structure

Cleaning ran before threading was joined onto the released files. Parents removed as spam, binary, PII, or excluded-group content are not present in the released records. A surviving reply whose parent was cleaned out is promoted to its own root. Depth and position may therefore have gaps. This is expected and documented, not a linker failure.

Threads are reconstructed from surviving posts. Messages removed during PII / spam / exclusion cleaning are not re-attached.

Crossposts

The corpus is organized one file per newsgroup. A message that was originally crossposted therefore appears once in each destination group’s file, with the same id.

Global audit of all *.threaded.jsonl.gz files:

Measure	Count
Total records	405,719,785
Unique `id` values	343,792,481
Duplicate records	61,927,304
Duplication rate	15.3%

Licensees who need a unique-message view should deduplicate on id. Hierarchy-level post and token totals above are record counts and include those copies.

Data Fields (Complete Schema)

Field	Type	Description
`id`	string	Anonymized Message-ID (`msg-<hex32>`)
`group`	string	Newsgroup name
`date`	string	Original Date header (when parseable)
`author`	string	Display name (email addresses redacted)
`subject`	string	Subject line
`text`	string	Cleaned post body
`thread_id`	string or null	Conversation identifier
`thread_position`	int or null	Position inside the thread
`thread_depth`	int or null	Depth from thread root

Repository Layout

One gzipped JSONL file per newsgroup:

usenet-alt-part1/alt.folklore.computers.threaded.jsonl.gz
usenet-alt-part2/alt.sysadmin.recovery.threaded.jsonl.gz
usenet-comp/comp.lang.c.threaded.jsonl.gz
usenet-humanities/humanities.lit.authors.shakespeare.threaded.jsonl.gz
usenet-sci/sci.physics.threaded.jsonl.gz
...

On part1 / part2: alt.* is split across two folders solely because the Hugging Face Hub caps a single directory at 10,000 files. The split is arbitrary and alphabetical — it carries no semantic meaning. part2 is not a different vintage, source, or quality tier. Treat the two folders as one hierarchy.

Data Quality & Processing

Three stages.

Stage 1 — Hierarchy-level exclusions

alt.binaries.* removed entirely (UUencoded binary content)
Adult-content newsgroups removed

Stage 2 — Record-level cleaning (identical to the unthreaded release)

Deduplication by Message-ID within each newsgroup file
Binary removal (UUencode and Base64 detection via MIME header inspection)
PII redaction — email addresses replaced with [email]
Message-ID anonymization — SHA-256 hashed to msg-<hex32>
Sensitive content filtering (spam, bulk PII patterns)

Stage 3 — Thread reconstruction

Header-only extraction from original MBOX zips
Graph construction via References
Deterministic thread_id, thread_position, and thread_depth
Join back onto the cleaned JSONL files

Retention versus the unthreaded corpus: 99.96% of posts, 99.97% of tokens.

Loading

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="OwnedByDanes/Usenet-Corpus-1980-2013-Threaded",
    filename="usenet-comp/comp.lang.c.threaded.jsonl.gz",
    repo_type="dataset",
)
df = pd.read_json(path, lines=True)

# Reconstruct a single thread
thread = df[df.thread_id == df.thread_id.iloc[0]].sort_values("thread_position")

Stream a hierarchy (alt.* needs both folders):

from datasets import load_dataset

ds = load_dataset(
    "OwnedByDanes/Usenet-Corpus-1980-2013-Threaded",
    data_files=["usenet-comp/*.threaded.jsonl.gz"],
    split="train",
    streaming=True,
)

Pull everything (bring a disk):

hf download OwnedByDanes/Usenet-Corpus-1980-2013-Threaded \
  --repo-type dataset --local-dir ./usenet-threaded

Deduplicate globally on id if you need unique messages:

# streaming example — keep first occurrence of each id
seen = set()
for row in ds:
    if row["id"] in seen:
        continue
    seen.add(row["id"])
    # use row

Intended Uses

Pre-training and continued pre-training of LLMs, under a licensing agreement
Dialogue and long-context modeling that benefits from explicit reply structure
Thread-level analysis, conversation modeling, and discourse research
Domain adaptation for technical, scientific, and historical internet discourse
Linguistic research and digital humanities
Debiasing and temporal language studies

Not intended for: re-identification of individuals, automated profiling, or spam generation.

Prior users of the unthreaded corpus report that comp.* and sci.* provide significantly cleaner signal than alt.*, and that temporal filtering (pre- vs post-1995) further improves output quality. Models trained on the related corpus: wyan/usenet-gemma-4-E2B-lora, wyan/usenet-gemma-4-E2B-lora-it.

Licensing & Access

Governed by LICENSE in this repository. Approving an access request does not grant a license. All use requires a separate written agreement.

Commercial AI training licenses — available, contact for pricing
Academic research licenses — available
Evaluation — use the free sample repository; no agreement needed
Redistribution — not permitted

Contact: usenetoverlord@gmail.com, or open a Discussion on the sample repository.

Provenance & Derivative Work

The underlying posts are public Usenet communications originally published in publicly accessible newsgroups. The compiled, cleaned, and threaded corpus is a derivative work produced through substantial original effort — format conversion, deduplication, sanitization, binary removal, PII redaction, Message-ID anonymization, and full conversation-graph reconstruction from Message-ID / References headers.

Citation

@dataset{usenet_corpus_1980_2013_threaded,
  author    = {OwnedByDanes},
  title     = {Usenet Corpus 1980--2013 (Threaded)},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/OwnedByDanes/Usenet-Corpus-1980-2013-Threaded}
}

Contact

📩 usenetoverlord@gmail.com

Downloads last month: 27

Total file size:

160 GB