han2ids v0.4

Read an IDS from a Mincho/ๅฎ‹ไฝ“ glyph image. v0.4 continues v0.3 on the corrected Noto/IPAmj/Jigmo em-box set. v0.1 is not used: 89.9% of the current test characters were in its training split.

Results

8,730 test rows, disjoint from training.

exact same tree valid
overall 84.54% 87.79% 100%
Noto Serif CJK JP 93.32%
IPAmj Mincho 84.63%
Jigmo 81.80%

v0.3 was 82.88% exact.

Architecture

35.2M parameters: ViT-S/16 encoder, four 384-wide decoder layers, 224px, 10,423-token character vocabulary, IDS operand-stack mask. Greedy decoding by default; --top-k returns ranked beam candidates.

Usage

pip install git+https://github.com/mkpoli/han2ids
huggingface-cli download mkpoli/han2ids-v0.4 han2ids-v0.4.pt --local-dir models
han2ids glyph.png --checkpoint models/han2ids-v0.4.pt

Training data

  • 159,521 images over 95,927 Han characters: Noto Serif CJK JP, IPAmj Mincho 006.01, Jigmo.
  • Common em-box renderer, font-aware gold from data/IDS.TXT, data/label-overrides.tsv and data/font-label-overrides.tsv.
  • Six epochs from v0.3; AdamW encoder 5e-5, decoder 2.5e-4, batch 32, bf16.

Limits

Clean em-box renders; crops, scans, Kai and sans are out of scope. IVD and regional variants are merged by default; strict classification is available in han2ids.status.classify.

Terms

Code MIT. Fonts and labels keep their own terms; see NOTICE.md and LICENSE.

Downloads last month
33
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support