mkpoli/han2ids-v0.1

han2ids v0.1

Read an ideographic description sequence (IDS) from a Mincho / 宋体 glyph image. This is the first checkpoint; v0.2 is the current model.

Scope is Mincho and Song serif faces only. Kai, handwriting and sans faces are out of scope.

Results

test	exact IDS	valid IDS
4,839 serif characters	80.10%	99.98%
8,730 mixed-font rows	74.27%	100%

By font on the mixed-font rows: Noto 92.15%, IPAmj 84.32%, Jigmo 63.47%.

Architecture

• 35.2M parameters, ViT-S/16 encoder and four transformer decoder layers.
• Character-level vocabulary of 10,402 tokens, maximum length 48.
• Greedy decoding with the IDS operand-stack mask.

han2ids-v0.1.pt is a state dict with itos and args; load it with han2ids.infer.Han2IDS.

Usage

pip install git+https://github.com/mkpoli/han2ids
huggingface-cli download mkpoli/han2ids-v0.1 han2ids-v0.1.pt --local-dir models
han2ids glyph.png --checkpoint models/han2ids-v0.1.pt

Training data

85,776 images from the serif compose export (Noto Serif CJK JP and Jigmo), eight epochs. See the han2ids repository for the training pipeline.

Limits

Clean em-box renders; tight crops, scans, Kai and sans faces are outside the training distribution. v0.1 is weaker than v0.2 on Jigmo; use v0.2 unless you need the earlier baseline.

Terms

Code MIT. Fonts and labels keep their own terms; see NOTICE.md and LICENSE. The checkpoint is derived from Noto Serif CJK JP (SIL OFL 1.1) and Jigmo (CC0 1.0).