Skip to content
Kotoshu Kotoshu 言修

Grammar models

Neural grammar correction, measured honestly

One model per language family — single-file ONNX, loaded through the gem, no Python runtime. Every score below was decoded from the exact artifact you download and scored by a pipeline built to reject inflated numbers.

English

kotoshu/grammar-en · 395 MB · 2026-11-19

CoNLL-14 (official M² scorer) 3.2× LanguageTool (0.168) on the same full set

0.535

422 ms/sent · 4 rnd · RSS 704 MB

French

kotoshu/grammar-fr · 271 MB · 2026-11-21

Held-out corruption benchmark

0.9803

Spanish

kotoshu/grammar-es · 184 MB · 2026-11-21

Held-out corruption benchmark

0.668

German

kotoshu/grammar-de · 220 MB · 2026-10-12

Measured on its held-out suite with an LLM-judge gate; see the model card.

F0.5 — card

Russian

kotoshu/grammar-ru · 213 MB · 2026-11-21

Held-out corruption benchmark

0.8333

Chinese · Japanese · Korean

kotoshu/grammar-cjk · 464 MB · 2026-11-17

Held-out + real-error benchmarks

0.661 / 0.788

Korean

kotoshu/grammar-ko · 356 MB · 2026-11-09

Held-out single-error benchmark 3.1× the grammar-cjk Korean weights (0.129) on the same 1,582-sentence slice

0.394

Urdu

kotoshu/grammar-ur · 184 MB · 2026-11-21

Held-out corruption benchmark

0.9137

446 ms/sent · 1 rnd · RSS 1061 MB

Portuguese

kotoshu/grammar-pt · 428 MB · 2026-11-19

Held-out corruption benchmark

0.571

Dutch

kotoshu/grammar-nl · 429 MB · 2026-11-19

Held-out corruption benchmark

0.565

Polish

kotoshu/grammar-pl · 427 MB · 2026-11-19

Held-out corruption benchmark

0.418

Italian

kotoshu/grammar-it · 478 MB · 2026-11-19

Held-out corruption benchmark

0.374

Ukrainian

kotoshu/grammar-uk · 427 MB · 2026-11-19

Held-out corruption benchmark

0.676

Turkish

kotoshu/grammar-tr · 428 MB · 2026-11-19

Held-out corruption benchmark

0.597

Serbian

kotoshu/grammar-sr · 426 MB · 2026-11-19

Held-out corruption benchmark

0.509

Greek

kotoshu/grammar-el · 421 MB · 2026-11-19

Held-out corruption benchmark

0.367

Suite capability by language. Spell = dictionary/frequency tier shipped; Rules = POS-keyed grammar rules in the engine.
Language Grammar correction Spell-check tier Grammar rules
English yes yes yes
French yes planned —
Spanish yes planned —
German yes planned —
Russian yes planned —
Chinese · Japanese · Korean yes planned —
Korean yes planned —
Urdu yes planned yes
Portuguese yes planned —
Dutch yes planned —
Polish yes planned —
Italian yes planned —
Ukrainian yes planned —
Turkish yes planned —
Serbian yes planned —
Greek yes planned —

F0.5 weights precision 5× over recall — a proofreading tool must not invent errors. Protocols: CoNLL-14 via the official M² scorer; fr/es via the strict span scorer on held-out corruption benchmarks with artifact guards.

In final evaluation

Five more languages at the gate

These models are trained and packaged; their scores are being measured now. Numbers appear when they survive the pipeline — not before.

The measurement law

Measured, not marketed

Every score on this page comes from the same frozen pipeline: decodes run against the exact artifact you download, and the scorer rejects the failure modes that inflate GEC numbers — output inflation, CJK tokenization mirages, span-size mirages. Where a comparison appears, the baseline was run on the same test set. Numbers that cannot survive this pipeline are not published.

Verified on Apple Silicon (arm64) and x86-64 Linux

Single-file ONNX per model, loaded through the gem with no Python runtime. Every shipped file passes a cross-architecture confidence check before publish.

gem install kotoshu
# fetch a grammar model into the gem cache
python scripts/fetch_grammar_model.py --lang fr   # → ~/.cache/kotoshu/grammar/fr