wime-syllable-mspy-25m

The 25M-parameter syllable model of wime, an Emacs Chinese input method. Given up to 64 characters of left context and a run of toneless syllables, it scores the Chinese characters that spell them. wime's decoder fuses it with a general language model (Qwen3.5-2B-Base) in a character-synchronous beam search.

• Input: each syllable is one token, written as its two Microsoft double-pinyin keys (wime-ms-pinyin-v1, composition encoding wime-composition-v2). Other layouts or full pinyin map to the same tokens syllable by syllable. Auxiliary (shape) codes are not part of the input.
• Architecture: an 8-layer decoder-only character LM (448 wide, 13,254 characters, 128-token window) from chinese-ime-lm; the model.py there loads this file.
• File: model.safetensors, FP16, sha256 2f8ab84a540431fda7f2a59cd59d0d00572b6a3a25a9fb1b91cc1bb4bde536a8. The embedded metadata carries the vocabulary, config and the full attribution chain; its license field is inherited from the base checkpoint.

Evaluation

Whole sentences on wime's sentence-trajectories-dev-v2 (798 cases judgeable by a person with the same left context), beam 32:

Setup	Correct
This model + Qwen3.5-2B-Base (weight 1.0)	722 (90.5%)
This model + Qwen3.5-0.8B-Base (weight 1.5)	706
This model alone	514

Training data

Fine-tuned in stages from a chinese-ime-lm checkpoint trained on the Chinese part of C4 (ODC-BY) and LCCC (MIT). Later stages, with automatic pinyin labels:

• LCCC dialogues (MIT); CHID (thu-coai)
• WildChat-1M (ODC-BY), Chinese Wikipedia 2026-10-01 (CC-BY-SA-4.0), Fineweb-Edu-Chinese-V2.1 (Apache-2.0 metadata; its card also requires the OpenCSG Community License), MNBVC Zhihu, forum, wiki and news parts (MIT card; upstream terms apply)
• Documentation: Python (PSF documentation license), Kubernetes, OpenHarmony, React and Vue (CC-BY-4.0), MDN (CC-BY-SA-2.5 prose), Apache Doris and PaddlePaddle (Apache-2.0), TRPL, Rust by Example and RIME (MIT)

The weights are released under CC-BY-SA-4.0, the most restrictive share-alike license among these sources; check the upstream terms above for your use.