han2ids v0.5

Read an ideographic description sequence (IDS) from a Mincho/宋体 glyph image. v0.5 continues v0.4 with labels that follow the printed form, ink-fitted renders and glyphs composed from two components.

Results

Held-out characters, character-macro agreement with BabelStone-derived references; 181 of the references were checked against the image by one reviewer. Full tables, intervals and limits are in benchmark/RESULTS.md.

Track v0.4 exact v0.5 exact v0.5 same-axis structure
Clean renders, 8,730 images of 4,752 characters 80.05% 83.77% 84.07%
Other styles, 1,346 images 87.15% 90.64% 90.64%
Reviewed sample, 181 images 82.67% 83.50% 83.50%

On the clean track the paired gain over v0.4 is 3.72 points (95% interval 3.09 to 4.31); on the reviewed sample the gain is within noise. v0.5 saw the clean track's framing in training, on training characters only.

A reading names the printed form. 䦳 (U+49B3) reads ⿵門鄉 in Noto Serif CJK JP, ⿵門鄕󠄂 (U+9115 U+E0102) in IPAmj Mincho and ⿵門鄕󠄁 (U+9115 U+E0101) in Jigmo; the unified reading of all three is ⿵門鄉. Only the 鄉 family has form labels so far.

Files

  • han2ids-v0.5.pt: the PyTorch checkpoint, a GlyphIDS state dict with itos and args. SHA-256 d5eba57784977a8e09384da6bba79dec1328bd44ea7a271af25dada39a90688d.
  • onnx/: split encoder and decoder in fp32 and int8, the vocabulary itos.json, the unified-forms table forms.json and a manifest with hashes. The int8 pair (40.2 MB) is what the browser demo runs.

Usage

pip install git+https://github.com/mkpoli/han2ids
hf download mkpoli/han2ids-v0.5 han2ids-v0.5.pt --local-dir models
han2ids glyph.png --checkpoint models/han2ids-v0.5.pt

Each line prints the reading, the unified reading and the confidence. --top-k 3 returns ranked beam candidates.

Architecture

35.2M parameters: ViT-S/16 encoder, four 384-wide decoder layers, 224 px input, 10,428-token vocabulary in which a character with a variation selector is one token, IDS operand-stack mask.

Training data

  • 159,521 images over 95,927 Han characters in Noto Serif CJK JP, IPAmj Mincho 006.01 and Jigmo, with test and validation characters held out.
  • Half of the rows drawn ink-fitted with a 6 to 30 px margin; a fifth of each batch composed as ⿰AB or ⿱AB from two glyphs of one face, never one whose unified reading is a validation or test label.
  • Three epochs from v0.4; AdamW encoder 1e-5, decoder 5e-5, batch 32, bf16, one seed.

Limits

Mincho and Song serif faces only; Kai, sans and handwriting are out of scope. The references are mostly unreviewed, and a regional sequence can differ from the form a font prints. Confidence is a geometric mean of token probabilities and is not calibrated.

Sources and terms

Labels: Andrew West, BabelStone IDS, 2025-06-27. Fonts: Noto CJK (SIL OFL 1.1), IPAmj Mincho (IPA Font License 1.0), Jigmo (CC0). Code and weights MIT; fonts and labels keep their own terms, see NOTICE.md.

Downloads last month
27
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support