MR3z4/persian-handwriting-ocr

Persian Handwriting OCR Dataset

Dataset Summary

A standardized dataset of Persian (Farsi) handwritten pages with word-level bounding-box annotations and transcriptions. The dataset is page-level: each sample is a full page scan; annotations are one row per word bbox on that page. This is the most flexible form -- users can train page-level OCR, word detection (DBNet/PaddleOCR), or derive word/line crops as needed.

• Pages: 1115 scanned pages (canonical IDs page_XXXX, under raw/pages/).
• Word annotations: 216202 rows (one per word bbox + transcription).
• Splits: page-level 70/15/15 train/validation/test (seed 42, no page leakage).
• License: CC-BY-NC-SA-4.0.

Format

Dataset/
  raw/pages/page_XXXX.jpg          # page scans (RGB JPEG, q95)
  data/train.csv                   # annotations for train pages
  data/validation.csv
  data/test.csv
  data/page_split_manifest.csv     # page_id -> split
  data/all_annotations.csv         # all pages before split
  data/all_annotations_normalized.csv
  reports/qc_report.md
  reports/class_balance_top100.csv
  reports/samples/page_*.png       # sample page overlays

CSV schema (one row per word):

file_name,text,text_raw,split,sample_id,page_id,block,line,x,y,x2,y2
raw/pages/page_0001.jpg,<normalized>,<raw>,train,page_0001_w_00000,page_0001,0,0,218,214,340,256

• file_name -- relative path to the page image (from Dataset/).
• text -- normalized transcription (see below); text_raw -- original.
• x,y,x2,y2 -- word bbox in page pixel coordinates (clamped to page bounds during crop, preserved here as-is).
• block/line -- IAM-style hierarchy.

Load with HuggingFace datasets:

from datasets import load_dataset
ds = load_dataset("csv", data_files={"train": "data/train.csv", "validation": "data/validation.csv", "test": "data/test.csv"})

For image models, join file_name to the page image on disk.

Collection Methodology

1115 pages of handwritten Persian notes (course material: transportation, statistics, mathematics, etc.) scanned at high resolution (A4@300dpi or phone photos) as RGB PNGs. Original scans under Raw Data/ are never modified; canonical copies under raw/pages/ are renamed page_XXXX.jpg (JPEG q95).

Annotation Process

Word-level bboxes + transcriptions, stored as block,line,x,y,x2,y2,text per word. Parsed with pandas (UTF-8), merged, normalized, and split by page.

Normalization Rules

text_raw is preserved; text is normalized:

1. Unicode NFC
2. hazm.Normalizer (remove_diacritics=True, persian_numbers=False):
   • Arabic ي/ك -> Persian ی/ک
   • Diacritics removed
   • ZWNJ kept (e.g. می‌شود)
   • Digits kept exactly as written (Persian digits stay Persian)
3. Whitespace collapse

Split Definition

• Page-level, random_state=42, 70/15/15 train/validation/test.
• No page appears in more than one split (verified in QC report).
• Mapping in data/page_split_manifest.csv.

Class Distribution

split	pages	rows
train	780	149506
validation	167	33946
test	168	32750
total	1115	216202

Top-100 token frequencies: reports/class_balance_top100.csv.

Intended Uses

• Page-level handwriting OCR / image-to-text (TrOCR-style).
• Word detection and recognition (DBNet, PaddleOCR, etc.).
• Deriving word/line crops for specialized training.
• Research on Persian script, mixed Persian/English, and math notation.

Limitations

• Mixed Persian prose, English tokens, and LaTeX-style math; some transcriptions are noisy.
• 30 empty-text rows were dropped; one 511-char LaTeX string is an anomaly (flagged in QC).
• No writer identity; style varies across pages.
• Bboxes are word-level; some formulas span multiple words.

Known GT Errors: Human Revision Required

The ground truth (GT) annotations contain errors that are already present in the dataset. Before using this dataset for training, evaluation, or any benchmark, it must be revised by human taggers. Known issues include:

• Invalid bounding boxes: 106 rows have bboxes that exceed page bounds (see reports/qc_report.md).
• Dropped empty-text rows: 30 rows with empty/whitespace-only text were removed during processing.
• Anomalous long transcriptions: 62 rows exceed 100 characters (mostly LaTeX-style math); one 511-char string is a known anomaly.
• Noisy transcriptions: some word transcriptions are noisy or incorrect.

The full audit is in reports/qc_report.md. Consumers should treat the GT as unverified and schedule a human review pass before relying on it.

Citation

@misc{persian_handwrite_ocr_2026,
  author       = {Mohammadzade, MohammadReza and Saemi, Payam},
  title        = {Persian Handwriting OCR Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  journal      = {Hugging Face Hub},
  howpublished = {\url{https://huggingface.co/datasets/MR3z4/persian-handwriting-ocr}},
  note         = {1115 pages, 216k word annotations with bboxes and transcriptions}
}

Contact

For questions or issues, contact Payam Saemi at payamsaemi@gmail.com or MohammadReza Mohammadzade at mohammadzade.m.r@gmail.com.