veda-ocr-rv-ms-noto-v2

Synthetic text-line images of Vedic Sanskrit in two accent-marking conventions, built on one shared text set so that only the notation differs. Used in Transfer Between Accent Notations in Vedic OCR: Learning Visual Conventions from Parallel Text (PACLIC 2026).

• text: accented ISO 15919 romanization (the target), identical for a key in both notations.
• image: the line rendered in Devanagari with Noto Sans Devanagari (size 32, 20 px padding), with accent marks placed in the RV or MS convention.
• accent_system: rv or ms. file_name: the image's original path.

split	rows		split	rows
rv_train	10,420		ms_train	10,420
rv_val	578		ms_val	578
rv_test	578		ms_test	578

The RV and MS splits contain the same texts; train, validation and test are disjoint by text.

Source text

The romanized texts are the Maitrāyaṇī Saṃhitā texts released with yzk/veda-ocr-ms (Tsukagoshi et al., 2025), keyed by Fushimi from von Schroeder's edition. The images of that dataset are not used; every image here is rendered from text.

Notation caveats

The "RV" view applies the Rigveda marking convention to these texts; it does not reproduce any specific printed RV edition, and RV numerals are not rendered. The MS grave accent is approximated with U+0952.

License

CC BY-NC-SA 4.0. The font, Noto Sans Devanagari, is licensed under the SIL Open Font License 1.1 (© The Noto Project Authors).

Citation

@inproceedings{TODO-paclic2026,
  title     = {Transfer Between Accent Notations in Vedic OCR: Learning Visual Conventions from Parallel Text},
  author    = {TODO},
  booktitle = {Proceedings of the 40th Pacific Asia Conference on Language, Information and Computation (PACLIC 2026)},
  year      = {2026}
}
@inproceedings{tsukagoshi-etal-2025-ocr,
  title     = {Towards Accent-Aware Vedic Sanskrit Optical Character Recognition Based on Transformer Models},
  author    = {Tsukagoshi, Yuzuki and Kuroiwa, Ryo and Ohmukai, Ikki},
  booktitle = {Proceedings of the Computational Sanskrit and Digital Humanities Workshop at the 19th World Sanskrit Conference},
  year      = {2025},
  pages     = {51--61},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2025.wsc-csdh.5/}
}