usernames

149,142,110 unique login-style usernames (1,713 MB of bytes) collected from public sources, filtered to [A-Za-z0-9._-]{3,64} with case kept, deduplicated exactly, and split by xxh3_64(name) % 64 == 0 into 2,328,816 held-out and 146,813,294 training names.

Files

• prep/names.parquet: every kept name with src (index of the source that first mentioned it, in the table order below), h (xxh3_64 of the name) and heldout; sorted by h.
• prep/train.bin, prep/heldout.bin: the same names in the same order, each followed by a zero byte (<eos>), ready to memory-map for byte-level training.
• prep/card.json: per-source counts, overlap between sources, length histogram, most frequent shapes.
• sources/<id>.parquet: the names extracted from each source before cross-source dedup, one name column, sorted by name.
• prep/names.parquet: the whole corpus, sorted by name: name, source (the owning source), heldout.

Curation

The corpus is curated: training.curate scores every name with a classifier fine-tuned from the teacher model on the judge labels below (expected quality 1–5, expected offensiveness 0–4, P(offensive ≥ 3)), and training.prep --curated keeps each name with probability (quality ramp 2.5 → 3.5) × (offensiveness flat to 0.1, ramp to 0 at 1.0), zero when P(offensive ≥ 3) > 0.2, decided by a hash of the name. Dropped 47,979,317 names: 27,418,319 at the quality floor, 3,718,801 at the offensiveness ceiling, 16,962,271 by the ramps' draw.

Labels

• labels/<set>.parquet: names scored by an LLM judge (training.judge --label): q is signup-form quality 1 (keyboard mash, ids) to 5 (a name a person clearly chose), o is offensiveness 0 (fine), 1 (edgy), 2 (crude), 3 (explicit or harassing), 4 (hateful or violent); a hash-strided sample of the corpus named after it. These train the curation classifier (training.curate).

Sources

id	what	raw	kept
xato	SecLists xato-net-10-million-usernames.txt, frequency-sorted	8,117,692	6,317,784
bsky	every Bluesky account's handle, first label, from plc.directory	44,847,753	34,729,153
stack	GitHub usernames keyed in bigcode/the-stack-username-to-repo	5,662,743	4,849,770
hn	Hacker News by from OpenPipe/hacker-news	617,993	484,921
reddit	Reddit comment authors, author column read over HTTP	13,930,423	10,772,830
urls	account segments of profile URLs in ks46/urls-sampled shards	471,414	390,827
roblox	Roblox usernames labelled normal	21,771	16,829
crates	crates.io users' GitHub logins from the database dump	61,307	50,995
kaggle	Meta Kaggle UserName; needs the landed Users.csv	30,213,080	26,729,645
gharchive	GH Archive hourly events since 2026-01-01: actor logins and repo owners	17,924,156	15,134,080
codeforces	codeforces.com user.ratedList, every rated handle	710,392	598,539
lichess	one month of Lichess rated games: White and Black players	1,737,598	1,507,796
jeanphorn	jeanphorn/wordlist usernames.txt	14,647	12,361
so	Stack Overflow Users.xml display names	3,814,430	2,814,763
se	every other Stack Exchange site's Users.xml display names	807,427	652,456
osm	OpenStreetMap changeset user names	1,298,150	1,088,543
enwiki	English Wikipedia contributor usernames, last revision per page	1,904,623	1,672,396
ssa	US Social Security first names 1880-2018, yob*.txt	42,329	41,950
wikis	contributor usernames of 15 large non-English wikis and Commons	1,127,276	896,302
sentiment140	Twitter handles in the Sentiment140 tweets	511,656	452,865
urls_full	profile handles of 980 hosts from the sorted ks46/urls dataset	41,828,518	23,180,585
atlas	profile handles of the same hosts from ks48/url-atlas's Common Crawl configs, newest first (partial: still being extracted)	21,456,049	16,746,720

raw is the count after the filter within the source; kept is after cross-source dedup, where the first source in this order owns a name. Stack Exchange display names of the form user1234567 are dropped.