Tool
Hangul share and mixed scripts
Paste or JSONL strings are counted as Hangul, Han, hiragana, katakana, Latin, Cyrillic, and Greek code points. The share uses that letter-like sum as the denominator. This is not a language-ID model. Nothing leaves this tab.
Script counts run in this tab. The paste is not uploaded.
What is counted
Each code point is placed in at most one letter bucket: Hangul (syllables, jamo, compatibility jamo), Han (BMP CJK), hiragana, katakana (including halfwidth), Latin (including fullwidth A–Z), Cyrillic, or Greek. Digits, punctuation, and emoji are not in the denominator.
Hangul share is Hangul count ÷ letter-like count. A missing share means the line had no letter-like code points. The default keep threshold is 30%, a form default, not a published Korean-data cut.
Lookalikes are a fixed list of Cyrillic and Greek letters that resemble Latin. Hangul Filler (U+3164 and jamo fillers) and fullwidth Latin are counted separately. None of these is a language label.
JSONL mode concatenates string fields on that line, then keeps or drops the whole object. kept.jsonl and dropped.jsonl are built here.
What this page does not do
- It is not language identification. English written in Hangul, or Korean written in Latin letters, is counted by script, not by meaning.
- CJK beyond the BMP, and scripts other than the seven buckets, are omitted from the share. They are not guessed.
- A keep/drop split is not a quality score. At most 20,000 lines or about two million characters.