Funovate-Technology/fontvibe-kaomoji

From FontVibe — browse and copy them at fontvibe.ai/tools/kaomoji. Mirror of Funovate-Technology/fontvibe-kaomoji (canonical copy, DOI 10.5281/zenodo.22945292); corrections go to the GitHub repo's issues. This mirror carries kaomoji.jsonl.gz (the full record set), kaomoji.csv, originals.json and stats.json; the other JSON cuts (core, extended, mixed, verbose) and the npm / pip packages are on GitHub. Paths below like data/… refer to the GitHub layout.

kaomoji

82,109 kaomoji (Japanese text emoticons), with semantic labels in 6 languages.

Browse and copy them at fontvibe.ai/tools/kaomoji — one click to copy, no install. The English page carries the 11,020 that render safely outside Japanese contexts; the Japanese page, 顔文字一覧, carries the whole corpus. This repo is the raw data behind both; every one of the 82,109 is in it.

Every publicly available kaomoji dataset we could find is covered here at 97.8–100%, and the few entries that aren't are ones we deliberately exclude. You don't have to take our word for any of it — a script checks it against the live upstream files.

82,109   kaomoji
69,679   with semantic labels (emotion / intent / subject) in en · ja · zh · es · pt · de
   106   original kaomoji, designed for feelings that had no kaomoji before
    52   categories
     4   tiers, split by how safely each renders outside Japanese contexts

Prefer a package to a download?

npm i kaomoji-dataset          # or: npx kaomoji-dataset 嬉しい -n 5

See npm/ for the JavaScript API.

---

Coverage

Merged from 7 independent source families, which turn out to barely overlap. Across the 21 pairs, the median Jaccard overlap is 0.3%; four pairs share not one entry; and exactly one pair clears 5% — rofimoji and emoji.nvim, the two English-side Linux pickers, which you would expect to. The kaomoji world is a set of near-disjoint islands, and that is why no single existing dataset is close to complete.

Every entry carries the sources it came from, so that is checkable in six lines:

import gzip, json, itertools, collections, statistics
by = collections.defaultdict(set)
for line in gzip.open("data/kaomoji.jsonl.gz", "rt", encoding="utf-8"):
    r = json.loads(line)
    for src in r["sources"]:
        by[src].add(r["text"])
o = [len(by[a] & by[b]) / len(by[a] | by[b]) for a, b in itertools.combinations(by, 2)]
print(statistics.median(o), max(o), sum(1 for x in o if x == 0))   # 0.003  0.079  4

Verify it yourself. The script downloads each upstream file at run time and diffs it against this dataset. Nothing is precomputed:

python scripts/verify_coverage.py     # stdlib only

Output as of the current release:

Source	Entries	Covered	Not covered
fdw/rofimoji — largest English-side set, 1.1k★	1,562	100.0%	0
Allaman/emoji.nvim	2,016	99.9%	3
aoguai/rime_kaomoji_dict	959	99.4%	6
mtripg6666tdr/Kaomoji_proj (MIT)	2,166	97.8%	48
Japanese IME dictionaries (kaomoji.com + Vector, 20 files)	77,086	100%	0
kaosute.net (web)	449	100%	0
FontVibe's own earlier library	1,525	100%	0

The bottom three are LZH/ZIP archives and a web page, not fetchable by a stdlib script; their numbers come from the same merge run and can be reproduced by re-running the collection pipeline. The four GitHub sources above are checked live, every run.

cspeterson/splatmoji is byte-identical to rofimoji's data upstream, so it isn't listed separately.

Why it isn't 100%

The 57 upstream entries not carried here are ones we don't consider kaomoji. The script prints all of them so you can disagree:

Excluded	Example
Emoji	🥺 🤯 🍱 — a different thing; there are better emoji datasets
Plain text with no face	RT OK カムサハムニダ 呼ばれて飛び出てジャジャジャジャーン
URLs	https://twitter.com/settings/applications/
Single characters and fragments	` · ﾟ · 【:εω
Japanese editorial annotations	(白目) (困惑) (適当) — stage directions, not faces
Decorative rules	☆⌒Ｙ⌒Ｙ⌒Ｙ⌒Ｙ⌒☆ · ━─━─━─━─━ — 飾り罫, used around kaomoji

Comparison is byte-exact, with three normalisations applied so that formatting differences upstream don't read as gaps: HTML entities decoded, scraped captions stripped ([噎住] ( *⊙~⊙), くコ:彡 sea), and whitespace ignored. The script reports both the raw byte-exact number and the normalised one.

How the line is drawn: by precedent, not by feel. Before excluding anything we check whether this dataset already carries entries of the same kind. That check moved 18 entries back in: bracketed single words like (白目) (17 precedents such as （大爆笑）), sparkle clusters like ꙳★*ﾟ (7 precedents), a face repeated in a row like (p.-)(p.-)(p.-) (10 precedents), and everything 3 characters long — 242 such entries are already here and all of them are real faces (・ω・ ◕‿◕ →_←), so length alone was never a reason.

Think something in that list is a real kaomoji? Open an issue — three rounds of exactly that feedback loop, run against ourselves, is how rofimoji went from 93.0% to 100.0%.

What's different

Ranked by how strong the evidence is. We separate what we measured from what we're guessing.

1. 106 original kaomoji — the part that didn't exist before

Not collected — designed, for feelings that had no kaomoji yet: 躺平 (lying flat), 社死 (social death), doomscrolling, saudade, Schadenfreude, mamihlapinatapai.

Marked "origin": "fontvibe-original". Checked for collisions three ways (byte-exact / whitespace-insensitive / NFKC) against every collected entry: zero.

ㅡ(ᴗ_ᴗ)ㅡ	lying flat — 躺平
(⊃///ω///⊂)	social death — 社死
ㅡ(￣ー￣)ㅡ✧	winning lying down — 躺赢
(・ω・)…(・ω・)ﾉ?	mamihlapinatapai — two people each waiting for the other to act

See data/originals.json.

2. Labels in six languages — measured

Every other public kaomoji dataset is English-only. rofimoji, the largest of them, carries one English keyword per entry (often literally one word).

Here, 69,679 entries carry emotion / intent / subject in en · ja · zh · es · pt · de, drawn from a controlled vocabulary of 275 values — 144 emotion, 75 intent, 65 subject — with the wording written to match how people in each market actually type online rather than translated word-for-word.

This matters because kaomoji search demand is mostly not English: Indonesia 15% · Mexico 11% · Brazil 10% · Chile 5% — over 40% non-English.

3. Semantic retrieval metadata — measured

Existing datasets give one keyword per entry, so searching happy recalls poorly. Three labelled axes plus 52 categories make retrieval substantially better.

12,546 entries (15%) are deliberately left unlabelled rather than guessed at. An honestly empty field beats a plausible wrong one.

4. Rendering metadata — a hypothesis, not a finding

ascii_safe, needs_cjk_font, display_width and the tier split are provided because no other dataset has them. Whether developers actually want them, we don't know — rofimoji has 1.1k stars with a single keyword per entry, which suggests existing needs are being met reasonably well. If you have an opinion, open an issue; that's the fastest way for us to find out.

---

Files

File	Entries	Size	Use it for
data/core.json	10,842	8.3 MB	Start here. Faces only, no Japanese dialogue — safest outside Japanese contexts
data/extended.json	168	0.2 MB	Nearly pure faces
data/mixed.json	29,209	25 MB	Face plus a short Japanese phrase
data/verbose.json	41,890	35 MB	The Japanese 顔文字＋セリフ tradition — face plus a full line of dialogue
data/kaomoji.jsonl.gz	82,109	4.7 MB	Everything, one JSON object per line
data/kaomoji.csv	82,109	6.6 MB	text,keywords, no header row — drop-in shape for rofimoji-style pickers
data/originals.json	106	0.1 MB	The original ones
data/stats.json	—	—	Counts, machine-readable

Tiers are assigned by how much of an entry consists of characters that appear in known pure-face samples. If you're building a picker for a non-Japanese audience, core.json is the subset you want.

Quick start

import json, gzip
rows = [json.loads(l) for l in gzip.open("data/kaomoji.jsonl.gz", "rt", encoding="utf-8")]

happy_es = [r["text"] for r in rows if "feliz" in r["keywords"].get("es", [])]   # 27,856
safe     = [r for r in rows if r["tier"] == "core" and not r["needs_cjk_font"]]

import rows from './data/core.json' with { type: 'json' }
const angry = rows.filter(r => r.keywords.ja?.includes('怒る'))   // 648 in core.json

Field definitions: SCHEMA.md.

---

Attribution

CC BY 4.0 asks you to credit the source "in any reasonable manner requested by the licensor". Here's what we ask for — copy whichever fits:

Markdown

Kaomoji data from [FontVibe](https://fontvibe.ai/tools/kaomoji) ·
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

HTML

Kaomoji data from <a href="https://fontvibe.ai/tools/kaomoji">FontVibe</a> ·
<a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a>

Plain text (for a CLI, a NOTICE file, an app's about screen)

Kaomoji data from FontVibe — https://fontvibe.ai/tools/kaomoji — CC BY 4.0

A few things worth saying plainly:

• Add rel="nofollow" if you prefer. It's your page; we're not asking you to pass ranking signals, and we'd rather you credit us with a nofollow link than not credit us at all.
• A link is not strictly required — CC BY's exact words are "to the extent reasonably practicable". In a terminal app or a printed piece, plain text is fine.
• Please link the site, not just this repo. fontvibe.ai/tools/kaomoji is the canonical home: it's where they're rendered and one-click copyable, which is what your users will actually find useful. The repo is the raw data behind it.

Citing it

There's a CITATION.cff, so GitHub's "Cite this repository" button works. BibTeX:

@misc{fontvibe_kaomoji,
  title        = {FontVibe Kaomoji Dataset},
  author       = {{FontVibe}},
  year         = {2026},
  version      = {1.0.6},
  howpublished = {\url{https://fontvibe.ai/tools/kaomoji}},
  note         = {82,109 kaomoji with semantic labels in six languages. CC BY 4.0}
}

License

• Data: CC BY 4.0. See NOTICE for upstream credits.
• The 106 original kaomoji and all metadata we produced (labels, names, tiers, rendering flags): CC0 — public domain, no attribution needed at all.

CC BY rather than CC0 for the whole set because one upstream source is MIT-licensed and MIT requires its copyright notice to be preserved, which CC0 cannot express.

Kaomoji themselves are a 25-year-old folk tradition with no identifiable authors. Nothing here is claimed as ours except the 106 marked fontvibe-original.

Contributing

Missing a source? Open an issue with a link — if it has entries we don't, we'll merge them and update the coverage table. Found a mislabelled entry? PRs welcome; id is stable.

---

Built and maintained by FontVibe. The browsable version is at fontvibe.ai/tools/kaomoji (English page: the 11,020 that render safely outside Japanese contexts) and 顔文字一覧 (Japanese page: the whole corpus).