From Newsgroup: news.software.nntp
On 10/6/26 01:06, Jesse Rehmer wrote:
I'm seeking cleanfeed.local entries, and bad_* files from news admins to filter as much as possible. I have 45 years of articles to filter, so
nothing
is irrelevant.
We took a different route.
Instead of collecting filter lists, we labelled the archive itself.
All numbers are from archive.usenet.blueworldhosting.com, read over
plain NNTP.
RESULT
A list of spam Message-IDs as INN history lines.
Load them into a fresh INN and it refuses those articles as duplicates. Cleanfeed never sees them.
There are two lists:
current 35.8M Message-IDs
core 27.5M, flagged by each of the last three passes
Each entry says what flagged it: label, NoCeM, rule or local model.
About 3% of the archive is flagged.
We estimate 1 in 22 flags is a real article.
The core list makes fewer mistakes.
This is a backlog cleanup.
99.6% of the flagged articles are older than 2025.
Most are from 1996-2010, about 25M.
The Google Groups wave adds 2.0M in 2023 and 0.6M in 2024.
2025 and 2026 together have about 140,000.
New spam is cleanfeed-ng's job.
SIZE
54,580 groups, 1.5 billion overview lines.
Crossposts count once, so about 1.15 billion articles.
A full pass reads all of OVER on 4 connections with COMPRESS DEFLATE,
capped at 500 Mbit/s.
It takes 2.5 to 3.5 hours.
606 GB of overview text came over the wire as 102 GB.
HOW
Labelling every article with a language model would cost about $20,000.
So we label a sample and train a small local model on it.
1. Sample.
One random window of articles per group, about 400,000 in all.
Claude Haiku labelled them from the headers: spam, ham or unsure.
The prompt says losing a real article is worse than letting spam
through.
2. Local model.
Logistic regression on the headers: Subject, From, groups,
crossposts, line count, reply or not.
It does 23,000 overview lines per second per core.
It calls spam only at 97% precision, ham at 98%.
Everything in between is let through.
3. The undecided.
Haiku labels them again, now with the first 600 bytes of the body.
One label per From+Subject+line count covers all copies.
This runs through the batch API at half price.
Batches of 100,000 came back in 10 to 17 minutes.
4. Other sources.
- NoCeM notices: 60.8M Message-IDs.
We skip four issuers.
Too much of what they listed was ordinary posts:
UltraHippo 1,077 of 1,523 checked (71%)
robot@pasdenom.info 27 of 28
nocem-fr@alphanet.ch 49 of 83
i2pn2-nocem@i2pn2.org 83 of 305 (27%, mostly de.*)
UltraHippo alone listed 44.8M of the 60.8M.
- Subject rules, not applied to replies or abuse groups.
- A From blacklist (Adam's, see below).
- Copy count: 20 or more copies of the same From+Subject+lines,
unless labelled ham.
14.8M keys are labelled: 11.3M ham, 2.5M spam, 1.0M unsure.
CHECKS
After each pass we pull a few thousand random flagged articles per
source.
Haiku labels them again with the body.
Where it disagrees, Claude Sonnet decides.
These articles are never used for training.
Share confirmed as spam:
flagged by first third now flags now
language-model labels 95.8% 78.2% 92.9% 7.3M
NoCeM 64.8% 94.2% 93.8% 4.5M
copy count 54.5% 85.4% 95.8% 2.5M
subject rules, From 90.3% 98.7% 98.7% 2.0M
local model 71.8% 87.8% ~93%* 19.5M
flagged in all 52.5M 48.1M 35.8M
estimated spam 37.0M 41.5M ~33.7M
estimated not spam 15.5M 6.0M ~1.6M
undecided, let through 164.0M 69.8M 17.8M
* not spot-checked yet, estimated
The third pass caught more spam, but 1 in 8 flags was a real article.
Most of those were job ads and classifieds in groups made for them.
Spam reports in news.admin.net-abuse.* got hit too.
Offers now count as spam only when posted where they don't belong.
COST
$378 in API fees, at list price, batch at half.
4.6 billion input tokens, 0.48 billion output, almost all Haiku.
The Sonnet checks are $9.94 of it.
Plus one machine and about 100 GB of NNTP per pass.
DISK
SQLite database 24 GB (NoCeM 9.3, flags 7.5, labels 5.2)
history, current list 35.8M lines, about 2 GB
LOOK AT IT
https://mbox-mo1.pugleaf.net:8182/
Read-only, no login.
- Groups: type a prefix, see the flagged count and share per group.
- Browse: enter a group name.
It shows the newest articles, marked "in filter", "unsure" or
"clean".
Click the article number to read one.
- Look up: paste a Message-ID to see if and why it is flagged.
- Labels and NoCeM: the raw labels and every NoCeM issuer.
Group and article pages are limited to 20 a minute.
Adam W. <
aw@somewhere.invalid> wrote:
a blacklist of Froms, all Thai subjects and froms, some keywords
in Subject related to Kindle, and one blacklisted URL in body.
We tested your rules against our labels.
- From list: 266 of your 624 addresses show up, on 1,847 articles.
96.6% of those were spam.
We added the list.
In the last pass it flagged 74,000 articles.
67 of 70 checked were spam.
- Kindle rule: 145 labelled, all spam.
Added too: 6,900 flagged, 6 of 6 checked were spam.
- 0xE0 can't be used archive-wide.
It is "a grave" in Latin-1 and starts the Indic scripts in UTF-8.
CODE AND DATA
code (MIT)
https://github.com/go-while/usenet-spamclassifier
data (CC0)
https://archive.org/details/usenet-spamclassifier-bwh-2026-11-09
The data is 7z-compressed TSV.
It has both lists, all six passes, the labels, the spot checks, NoCeM,
the history file, the models and the full database.
No article bodies.
archive.org also offers it as a torrent.
Work in progress: data upload to archive.
--
.......
Billy G. (go-while)
https://pugleaf.net
@Newsgroup: rocksolid.nodes.help
irc.pugleaf.net:6697 (SSL) #lounge
TOR-IRC: ij7fmstcudrwyty5p6iloodeenur4wksovlhbyrfubat7eixplmrxvqd.onion:6667 --- Synchronet 3.22a-Linux NewsLink 1.2