veda-ocr / README.md
yzk's picture
v1.3.1: printed page of 55 MS lines (veda-ocr-datasets 1b9319f)
d47c145 verified
|
Raw History Blame Contribute Delete
8.63 kB
metadata
pretty_name: Veda OCR Dataset
license: cc-by-nc-4.0
language:
  - sa
task_categories:
  - image-to-text
dataset_info:
  - config_name: ms
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
      - name: id
        dtype: string
      - name: work
        dtype: string
      - name: image_source
        dtype: string
      - name: text_source
        dtype: string
      - name: volume
        dtype: int64
      - name: page
        dtype: int64
      - name: line
        dtype: int64
  - config_name: ks
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
      - name: id
        dtype: string
      - name: work
        dtype: string
      - name: image_source
        dtype: string
      - name: text_source
        dtype: string
      - name: volume
        dtype: int64
      - name: page
        dtype: int64
      - name: line
        dtype: int64
  - config_name: ts
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
      - name: id
        dtype: string
      - name: work
        dtype: string
      - name: image_source
        dtype: string
      - name: text_source
        dtype: string
      - name: volume
        dtype: int64
      - name: page
        dtype: int64
      - name: line
        dtype: int64
configs:
  - config_name: ms
    data_files:
      - split: train
        path: data/ms/train-*.parquet
  - config_name: ks
    data_files:
      - split: train
        path: data/ks/train-*.parquet
  - config_name: ts
    data_files:
      - split: train
        path: data/ts/train-*.parquet

Veda OCR Dataset

Line images from printed editions of Vedic Saṃhitās, paired with transcriptions (ISO 15919, NFC). One config per text. There is no fixed train/test split, so choose your own; splitting by page avoids leakage between neighbouring lines.

Config Text Printed edition Rows
ms Maitrāyaṇī Saṃhitā von Schroeder, 1881–1886, vols. 1–4 12,396
ks Kāṭhaka Saṃhitā von Schroeder, 1900–1910, vols. 1–3 12,493
ts Taittirīya Saṃhitā Weber, 1871–1872, vols. 1–2 14,947
from datasets import load_dataset

ds = load_dataset("yzk/veda-ocr", "ks")

Columns

Column Meaning
image One printed body line, cropped and cleaned of neighbouring lines and margin material
text Transcription of the line
id Line id, unique within the config
work Config name (ms, ks, ts)
image_source Printed edition the image comes from
text_source Source of the transcription
volume, page, line Location in the printed edition

Line ids in ms

Most ms ids are <volume>-<page>-<n>, where n (also the line column) numbers the line strips cut from that page by the original segmentation. Strips include page headers and other non-body lines, so n is not the printed line number.

Some lines added in v1.3 had no strip of their own. Either the segmentation had merged them into the gap between two strips, and they were re-cut from those two neighbouring strips, or no usable strip existed, and they were cut afresh from the Bayerische Staatsbibliothek page images (see the image sources below). Lacking a strip number, they have ids <volume>-<page>-r<n>, and n (also the line column) counts the body text lines of the page, as the printed margin numbers do. The two numberings therefore differ: 3-145-r4 is the fourth body line of page 145, while 3-145-4 is the page's fourth strip.

The page column always gives the printed page. For 55 lines of volume 4 the page in the id is not the printed page (ids were kept when the page was corrected in v1.3.1): 4-69-* lines are on page 71, and 4-132-*, 4-133-*, 4-134-* lines are on pages 130, 131, 132.

Label conventions

  • Printed section numbers are kept in the image and written as // n // in the label.
  • Footnote markers may remain in the image but are not in the label.
  • A word broken at the line end is marked with - in the label.
  • Punctuation (/, //) follows the transcription, not always the print.

Known limitations

  • Some images keep tiny specks of neighbouring-line strokes.
  • Punctuation is not yet aligned to the print.
  • Lines added in v1.3 were accepted by an automatic OCR check and spot-checked, not reviewed one by one.
  • Since v1.3 nearly all ms images are grayscale crops of the Bayerische Staatsbibliothek page images, like the ks and ts images. 3 ms lines whose position on the page could not be confirmed automatically keep their earlier image: a binarized crop from the original line strips, smaller and cut more tightly (accents above the line may be clipped).

Sources

Transcriptions

The labels come from the electronic texts of the Maitrāyaṇī, Kāṭhaka and Taittirīya Saṃhitās prepared by Makoto Fushimi (Ōsaka). They were converted to ISO 15919, NFC-normalised, split into printed lines and corrected while building this dataset. Versions of these e-texts are also available from TITUS, for anyone who wants the continuous text.

Images

Config Printed edition Digital copy
ms Leopold von Schroeder (ed.), Maitrāyaṇī Saṃhitā, 4 vols., Leipzig: F. A. Brockhaus, 1881–1886 Bayerische Staatsbibliothek München, bsb11542039–bsb11542042 (MDZ)
ks Leopold von Schroeder (ed.), Kāṭhakam. Die Saṃhitā der Kaṭha-Çākhā, Leipzig: F. A. Brockhaus, 1900–1910 Vol. 1: Internet Archive tgQv_kathakam-die-samhita-der-katha-cakha-by-leopold-von-schroeder-1900-anon; vols. 2–3: Internet Archive in.ernet.dli.2015.487463 (Digital Library of India)
ts Albrecht Weber (ed.), Die Taittirîya-Saṃhitâ, Indische Studien 11–12, Leipzig: F. A. Brockhaus, 1871–1872 Bayerische Staatsbibliothek München, bsb10996666, bsb10996667 (MDZ)

The printed editions are in the public domain. The Bayerische Staatsbibliothek digital copies carry the rights statement No Copyright – Non-Commercial Use Only. This is one reason the dataset is released for non-commercial use only.

Changelog

  • v1.3.1 (2026-11-15): Corrected the page value of 55 ms lines in volume 4 that were attributed to the wrong printed page: the lines with ids 4-69-* are on page 71 and those with ids 4-132-* to 4-134-* are on pages 130–132 (ids are unchanged; see "Line ids in ms"). The build manifest now records whether the code that built it had uncommitted changes. No line text or image changed.
  • v1.3 (2026-11-14): 1,639 printed body lines that the earlier crops had lost were added (ks 962, ms 677). KS lines were re-cropped from the page scans at positions fixed by their neighbours; MS lines came from unused line strips, from the seam between two overlapping strips, or (511 lines) from the Bayerische Staatsbibliothek page images. A line was added only if an OCR reading of the cleaned image matched its transcription (at most 10% character errors, matching line starts and ends), and lines whose printed section number or line-end hyphen could not be read from the image were left out, as were lines whose transcription nearly repeats another line of the same page. MS line images were re-cut in grayscale from the Bayerische Staatsbibliothek page images, about twice as large as before and with room for the accents above and below the line (12,393 of 12,396 lines; each new crop was matched against the old image or, where the old image was the noisier one, read by OCR against the transcription or checked by eye; the 3 lines that could not be confirmed keep their old binarized image). After a review of flagged lines, 67 labels were corrected and 9 lines removed.
  • v1.2 (2026-11-13): 187 labels corrected and 8 line images re-cropped. Candidates came from out-of-fold OCR predictions and pattern checks (leaked markup, duplicated text, line-boundary slips, private-use characters); each change was checked against the image by several models, with the uncertain ones decided by a human.
  • v1.1 (2026-11-12): first public release.

How to cite

@misc{veda_ocr_dataset,
  title  = {Veda OCR Dataset},
  author = {Tsukagoshi, Yuzuki},
  year   = {2026},
  url    = {https://huggingface.co/datasets/yzk/veda-ocr}
}

License

CC BY-NC 4.0. When you use the data, please also credit the printed editions, the digital copies (for the MS and TS images, "München, Bayerische Staatsbibliothek") and Makoto Fushimi's e-texts listed above.