Himalaya OCR Training (himalaya-ocr-training)

Page-level OCR training pairs for Nepali document OCR: every PDF page of 1,000 books from the OLE Nepal E-Pustakalaya repository (pustakalaya.org), rendered to an JPEG image and paired with the OCR text produced by Surya v0.17.1 (multilingual auto-detect).

Field	Value
Books / pages	1,000 books, 110,796 paired pages
Image format	JPEG q85, rendered at 150 DPI (same render path the OCR run used: PyMuPDF, LANCZOS, ≤2400 px width)
Text labels	OCR output normalized (NFC, control-character stripped, ZWJ/ZWNJ/danda preserved, mojibake filtered)
Avg OCR confidence	0.922 (line-averaged per page)
QC flags	725 empty-text pages, 3,200 pages flagged for review

Structure

images/<book-uuid>/page_NNNN.jpg    # one JPEG per page
manifest.jsonl                      # one record per page (see below)
dataset_summary.json                # build statistics + provenance

Manifest record

{"id": "<uuid>__page_0001",
 "uuid": "<book-uuid>", "page": 1,
 "image": "images/<uuid>/page_0001.jpg",   # path relative to this root
 "text": "<normalized OCR text>",
 "confidence": 0.88,                        # page-level average
 "width": 1240, "height": 1755,             # image dimensions in pixels
 "qc": {
   "empty": false,            # no text was extracted
   "needs_review": false,     # replacement chars or assembly review flag
   "preeti_suspect": false,   # heuristic: residual legacy-font gibberish
   "replacement_chars": 0},
 "meta": {
   "title": "...", "author": "...", "publisher": "...",
   "language": "नेपाली",
   "detail_url": "https://www.pustakalaya.org/documents/detail/<uuid>/",
   "license": "CC BY-NC-ND 3.0 ..."}}

All pages are kept (with QC flags); filter at training time — e.g. skip records with qc.empty, qc.needs_review, or low confidence.

Loading

import datasets as hfds

ds = hfds.load_dataset("edumaharshi/himalaya-ocr-training")
ds = ds.filter(lambda r: not (r["qc"]["empty"] if isinstance(r["qc"], dict) else r["qc"]["empty"]))

Provenance note

Labels are self-generated OCR (Surya), not human transcription — they are "pseudo-ground-truth": good for sequence-level ICR/OCR pre-training and self-supervised line models, not a substitute for hand-verified evaluation data.

License

Source material: OLE Nepal E-Pustakalaya, licensed CC BY-NC-ND 3.0 — attribution required, non-commercial use, no derivatives. This dataset is uploaded privately for research/learning; do not redistribute it publicly or use it commercially.