cursive_v2 / README.md
mic7ch's picture
Reupload as parquet (1940 rows); drop imagefolder layout
5c0837b verified
metadata
license: cc-by-4.0
language:
  - mnc
tags:
  - ocr
  - htr
  - manchu
  - cursive
pretty_name: Cursive Manchu (v2)
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet

Cursive Manchu — v2

Word-level crops from the Manchu manuscript Talking Dead (preface + chapter 01), paired with Möllendorff romanization and Manchu Unicode (Mongolian block).

Splits

  • test — 1,940 word images from the preface (14 pages) and chapter 01 (12 pages).

Columns

  • image — PNG crop of a single Manchu word in cursive script, rotated 90° counter-clockwise and padded to 480×64 white background.
  • roman — Möllendorff romanization.
  • manchu — Manchu Unicode (renders correctly with Noto Sans Mongolian or any Manchu/Mongolian-capable font).

Usage

from datasets import load_dataset
ds = load_dataset("mic7ch/cursive_v2", split="test")

Notes

  • Manchu codepoints use the Mongolian block (U+1820–U+1876). Some are formally named "MONGOLIAN LETTER SIBE …" in Unicode — Sibe and Manchu share the script.
  • Romanizations were checked against Jerry Norman's A Comprehensive Manchu–English Dictionary (2013). A residual review set of ~80 romans (proper nouns, loans, inflected forms, possible transcription typos) was inspected manually; see repository history.
  • Companion to mic7ch/cursive_v1 (327 word crops from a 1680 Manchu military document).