pathnirvana/sinhala-pali-ocr

Sinhala and Pali OCR: old prints

These are scanned pages of Sinhala and Pali books, mostly Buddhist texts printed between the 1870s and 2000s, each paired with a corrected transcription. They were collected to fine-tune a page-level OCR model (e.g. LightOnOCR-2) for old Sinhala-script print, made by Path Nirvana.

Target format

Each target is the page text as UTF-8 Unicode Sinhala (NFC):

• Lines: one printed line becomes one line of text. A word split by the printer across a line end keeps its hyphen (word-\n).
• Paragraphs: a blank line separates paragraphs, headings, verses and footnotes.
• Bold: printed bold is marked **bold**.
• Footnotes: reference numbers are superscript digits (¹ ² ³).
• Left out: running headers, page numbers and printer's footer codes.

The prompt column says when a target falls short of this format. The image is then given to the model with that prompt as a text instruction, so that pages with no prompt keep the full format:

prompt	meaning
null	full format
nobold	the source transcription has no bold markup (the acts pages)

crop is [x0, y0, x1, y1] to apply to the image before use, or null for the full page. A crop is used where the page holds text that the transcription lacks.

Files

The data is in Parquet shards, data/{train,validation,test}-*.parquet, with one row per page. The columns are id, image (the original JPEG bytes), crop, text, prompt, source, stratum, lang and year.

from datasets import load_dataset
ds = load_dataset("pathnirvana/sinhala-pali-ocr", split="train")
row = ds[0]
img = row["image"].crop(row["crop"]) if row["crop"] else row["image"]

Sources

source	pages	images	text
bjt	3,301	Buddha Jayanti Tripitaka (Sri Lanka government edition, 1960s–80s), Pali and facing Sinhala translation	typed text from tipitaka.lk. Printed line breaks were recovered by aligning Tesseract line boxes. 30 pages per volume and language
pitaka-books	4,795	old scanned prints in the tipitaka.lk library	corrected editions from pitaka.lk/books, cut into pages by alignment. At most 150 pages per print, from 32 books
acts	410	Sri Lankan acts, 1981–1989	the pre-2000 pages of avishadilhara/sinhala-ocr-lk-acts-1010 (CC BY 4.0), with headers removed
opus-lib	92	single pages of library books printed 1870s–1980s	transcribed by Claude Opus. Not human-checked, so expect a few % character errors and imperfect bold

The page transcriptions come from corrected editions, which can differ from the printed page in places: spelling, punctuation, small edits. Expect a floor of about 1–2% character differences.

Splits

• test: whole books never trained on. These are BJT volumes 12, 29 and 47, and the pitaka.lk books suvisi-gunaya (1964 print) and chathurarya-sathya (2007).
• test, other sources: the acts follow the source dataset's split, and 10% of the opus-lib pages are test.
• validation: a small set of pages from all sources, used to track loss.

Credits

The acts pages come from Sinhala OCR – Sri Lankan Acts Dataset by avishadilhara (CC BY 4.0). The BJT text is from tipitaka.lk, and the book texts and scans are from pitaka.lk and the tipitaka.lk library.