dadeum-gec-base

A small Korean spelling, spacing and grammar corrector that fixes only what is wrong and leaves the rest of the sentence alone. It comes from dadeum, a grammar checker that runs on the user's own Mac, and is trained only on openly licensed text: no training pair was annotated by an AI provider.

틀린 곳만 고치고 나머지는 그대로 두는 작은 한국어 맞춤법 · 띄어쓰기 · 문법 교정 모델입니다. 기기 안에서 도는 맞춤법 검사기 다듬에서 나왔고, 공개 라이선스 글과 정답만으로 학습했습니다(AI 가 단 정답은 쓰지 않았습니다). 앱이 쓰는 규칙 엔진보다는 낮고, 아래 비교표는 다듬에 유리한 우리 평가셋에서 잰 것입니다 (Limitations 참고).

The model rewrites a sentence; each change it makes is then judged by its own log-likelihood margin and kept only at or above the margin in gating.json, and guards drop changes that touch non-Hangul, non-ASCII characters, add or remove two or more characters, change a number, or jump three or more jamo to a word absent from the open word list (known_words.bin). Use dadeum_correct.py, which does all of this; plain generate() skips the judgement and the guards.

pip install torch transformers
python dadeum_correct.py "오늘은 날씨가 좋아서 산책을 갔는데 사람이 많앗다."

from dadeum_correct import Corrector
corrector = Corrector(".")
corrector.correct("회의는 다음주 화요일에 하기로 했어요.")
corrector.edits("...")   # [(start, end, replacement), ...]

Model

BART, 6 encoder and 6 decoder layers, width 768, 123.9M parameters, started from KoBART (gogamza/kobart-base-v2).

Training data

Openly licensed text and targets only - no pair was annotated by an AI provider.

Pairs	Source text	Target
Rule-engine corrections	nsmc (CC0), unlabeled kocohub comments, Korean Wikipedia and Wikinews (CC BY-SA 4.0), 정책브리핑 (KOGL Type 1)	the dadeum rule engine, with its parts fitted to AI-annotated data switched off
Wikipedia editors' corrections	Korean Wikipedia page history (CC BY-SA 4.0)	the editor's corrected sentence, filtered (no evaluation sentences, no content swaps, no foreign-script or number changes)

Training steps, from the base: clean_full (legacy_clean:60, wiki_edits:7)

Evaluation

Measured against gold that AI annotators wrote (not distributed); F0.5 over character edits, touching edits merged. That gold trained nothing, but its val split chose the training checkpoint and the margin in gating.json, so the val scores are slightly optimistic.

Measure	Value
Blind half of val (1,196 sentences), F0.5 (P / R)	.890 (.964 / .682); the dadeum rule engine: .899 (.946 / .750)
Development, 300 sentences, F0.5	.918; rule engine .950
In-domain test, 350, F0.5	.904; rule engine .927
Sealed domains (reviews, comments, policy, requests), 870, F0.5	.820; rule engine .863
Confirmation domains, 600, F0.5	.803; rule engine .831
News, business and comments (single annotator), 799, F0.5	.788; rule engine .820
Kor-Lang8 rebuild (learners), 2,000, F0.5	.089; rule engine .088
Correct workplace email (27 sentences and the whole email), sentences changed	0
Clean Wikipedia sentences (19,759), sentences changed	694, 602 of them spacing only; rule engine 424
CPU latency per sentence, Apple M4 Pro, 4 threads	120 ms median, 325 ms p95

Next to other Korean correctors

The same sentences and scorer, F0.5. This model runs at its margin; the other open correctors as their cards say; the LLMs with the same annotation guideline in the prompt. The gold rewards minimal edits; on learners' writing (Kor-Lang8), where the reference corrections rewrite more, KoBART gec_kr leads. No public Korean leaderboard for this task was found, so this is our table, not a third-party ranking.

System	Sealed domains (870)	Confirmation domains (600)	Kor-Lang8 (2,000)
this model (dadeum-gec-base)	.820	.803	.089
dadeum rule engine (the app)	.863	.831	.088
Claude Sonnet 5.5 with an annotation guideline	.895	.887	not run
Qwen3-8B	.548	.562	.127
KoGrammar distil (theSOL1)	.481	.435	.107
KoGrammar base (theSOL1)	.476	.439	.104
ET5 typos corrector (j5ng)	.408	.347	.102
KoBART gec_kr (Soyoung97)	.216	.187	.266

Limitations

• The benchmark above is ours and favours this model: its gold was written by Claude under our annotation guideline, and it rewards minimal edits, so correctors that rewrite score low on it. No human checked that gold.
• It learns from the dadeum rule engine's corrections and does not beat that engine, which the dadeum app uses. Its point is an open license and a single file that runs on a CPU.
• It is a corrector for native writing. On learners' writing corrected by people (Kor-Lang8), where the references rewrite whole phrases, it rarely changes anything and KoBART gec_kr does far better.
• It shares the rule engine's blind spots; it corrects spelling, spacing and particles far better than word choice or style.
• One sentence at a time, up to 128 tokens; longer text is split at sentence ends.

License

CC BY-SA 4.0 (LICENSE); attributions in NOTICE. dadeum_correct.py is MIT, like the dadeum repository.