grammar-ko (v10)

Korean grammar correction — GECToR-style token tagger, single-file ONNX, no Python runtime.

Measured, not marketed. Held-out slice (1,582 sentences, strict span scorer): F0.5 0.394 (precision 0.61, recall 0.16), decoded from this exact artifact at margin gate 1.0, rounds 4. Baseline comparison — the previous published Korean weights (grammar-cjk ko cell) score 0.114 on the same slice: 3.5× on identical measurement.

Trained on 359,212 balanced corruption pairs (josa particle swaps, attested verb/adjective conjugation swaps, homophone confusables — uniformly drawn, attested-form law) over a Korean Wikipedia corpus.

Files

• gec-tagger.int8.onnx — the model (load through the kotoshu gem)
• labels.json, pieces.json, charsmap.json, tokenizer.json, tokenizer_config.json, keep_ids.json — serving vocabulary

License BSD-2-Clause.