Datasets:
Born-digital Indic rows: text can be recovered exactly without OCR (open tool, measured)
Many Indian-language rows come from born-digital PDFs (often Word exports) whose text layer is garbled, rather than from scans. That text can be recovered from the PDF itself, without OCR and without the original font, with an open tool, Unshape (v0.3.0).
Measured so far:
4 hin_Deva rows with invalid words (first lines read from the page images before decoding, 119 words): stored text 58.8% exact words → Unshape without any font 92.4% → with the installed font 97.5%.
Fresh pre-registered Hindi Word files (Mangal, Kokila, 592 words each): 97.8% / 97.3% exact words, vs Tesseract 5.3.4 95.8% / 96.3% and the PyMuPDF text layer 70.6% / 69.3%.
In a random sample of OCR-path rows across 6 Indic configs (90 checked), about 3 in 10 were born-digital rather than scans (rough estimate). About half of those use Unicode fonts and can be decoded on a CPU instead of being OCR’d; the rest use older legacy fonts that Unshape cannot decode yet.
Scope: proven for Hindi Word PDFs; Tamil and Bengali promising on development files; scans still need OCR.
Details and evidence: huggingface/finepdfs#4
Code and every test (including the ones below their bar): https://github.com/novain-shibu/unshape (DOI 10.5281/zenodo.23479502)
Happy to run it over the born-digital Indic rows and share recovered text with per-row confidence, if that would help.
(Written with AI assistance; I checked the numbers and wording.)