deberta-v2-tiny-japanese-ime-web

deberta-v2-tiny-japanese-ime-web

ku-nlp/deberta-v2-tiny-japanese をブラウザーで有限候補の文脈比較に使うため、 ONNX opset 18 へ変換し、動的 q8 量子化した配布物です。追加学習済み IME 生成モデルではありません。 公開デモでは、このモデルが候補外の日本語を生成することはなく、外部サービスから取得した有限候補の 変化部分だけを周辺文脈と比較します。

Source and attribution

• Source model: ku-nlp/deberta-v2-tiny-japanese
• Pinned source revision: 0427645e8cf44ee83b4a0b5f4498274d89e02adb
• Source/model license: CC BY-SA 4.0
• Modification: ONNX opset 18 export followed by ONNX Runtime dynamic q8 quantization (--avx2)
• Transformers.js runtime tested: 4.2.0

The original model authors and KU-NLP retain attribution for the source model. This derivative is distributed under the same CC BY-SA 4.0 license. See the source model card for training data, limitations, and primary attribution.

Intended use

This artifact is intended for masked-language-model scoring in modern browsers. The accompanying demo uses pseudo-log-likelihood over a finite changed span. It is not a standalone Japanese IME, does not transliterate romaji, and should not be used to claim measured typo accuracy without a separate evaluation.

Files and integrity

The receipt in EXPORT_RECEIPT.md records the reproducible export commands and local runtime probe.