loohhoo/detangutify-data

Detangutify Tangut Recognition Data

Data files used by Detangutify, an image and drawing recognizer for Unicode Tangut characters.

Template archive

templates_aug.npz contains the template bank used by the current recognizer:

• Unicode range: U+17000–U+187F7 (6,136 characters)
• Feature matrix: 325,152 × 1,835 (float32)
• Labels: Unicode codepoints stored in y
• Normalization arrays: mean and std
• Provenance arrays: source_id and source_names

Source	Template rows
Legacy Detangutify templates	196,352
Noto Serif Tangut	49,088
Tangut Yinchuan	49,088
BabelStone Tangut Wenhai	24,488
GlyphWiki KAGE/Mincho	6,136

The GlyphWiki rows contain one unmodified template for every character in the range. They were generated from GlyphWiki's official bulk KAGE dump with exact component dependencies and historical component revisions preserved. Font and GlyphWiki images were processed through the same crop, center, binarization, stroke-normalization, and 64×64 feature pipeline.

Usage

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(
    repo_id="loohhoo/detangutify-data",
    repo_type="dataset",
    filename="templates_aug.npz",
)

data = np.load(path)
X = data["X"]
y = data["y"]
mean = data["mean"]
std = data["std"]
source_id = data["source_id"]
source_names = data["source_names"]

Each row of X is a recognition template and the corresponding value in y is its Unicode codepoint. Recognition results should be ranked by character, keeping the best score across that character's templates.

Provenance

This repository began with the data published in raycosine/detangutify-data and has been expanded with additional Tangut font forms and GlyphWiki renderings. The archive contains extracted feature vectors rather than the original font files or GlyphWiki image cache.

This is recognition support data, not an end-to-end neural OCR model. It is intended for Tangut OCR experimentation, search, and document-processing tools.