jai9999/hindi-english-line-ocr

Hindi-English Line OCR Dataset

Printed Hindi, English, and mixed-script line images from scanned documents, rendered PDFs, and synthetic text. Images are embedded directly in Parquet format with exact Unicode transcriptions. This final release contains 98,195 verified line crops organized into standard training, validation, and test splits.

Splits and Script Distribution

Split	Real	Synthetic	Total
train	8,833	83,000	91,833
validation	1,224	2,000	3,224
test	1,638	1,500	3,138
Total	11,695	86,500	98,195

Real Images

Split	Hindi	English	Mixed	Digits / Symbols / Blank	Total
train	6,450	1,597	249	537	8,833
validation	864	171	155	34	1,224
test	1,109	194	255	80	1,638

Synthetic Images

Split	Hindi	English	Mixed	Digits / Symbols / Blank	Total
train	34,521	12,267	35,695	517	83,000
validation	704	302	994	0	2,000
test	347	275	878	0	1,500

Script categories inspect the actual transcription characters. Hindi indicates Devanagari letters are present; English indicates ASCII Latin letters are present; mixed requires both scripts. Digits, punctuation, symbols, and blank crops are reported separately.

Provenance and Sources

The dataset balances authentic printed historical and modern publications with diverse synthetic rendering:

• Real document collections:
  • Government and official publications: Central Hindi Directorate programmes and guidelines, ISRO space science glossary, Ministry of Earth Sciences administrative glossary, Rajbhasha Bharti, and NHAI application forms.
  • Educational textbooks: NCERT secondary science textbooks and Hindi literature collections (Kshitij).
  • Historical Hindi literature: Chandrakanta, Skandgupt, Mansarovar, Bhasha Itihas, Bhoo Vigyan, Bandi Jeevan, Bharat Vibhajan, Volga Se Ganga, Gunahon Ka Devta, Ramcharitmanas, and Chintamani.
  • Classical English literature: Pride and Prejudice, Great Expectations, and A Tale of Two Cities.
  • IndicVisionBench Devanagari line corpus.
• Synthetic generation:
  • Text sourced from Hindi Wikipedia, literary prose, and bilingual term pairs.
  • Rendered across open-source Devanagari and Latin fonts with varied weights, sizes, background textures, blur, and aspect ratios.

Methodology and Quality Verification

1. Rendering and line segmentation: Scanned books and PDFs were rendered at 300 DPI. Line bounding boxes were extracted using automated text detection.
2. Transcription and agreement filtering: Real lines were transcribed with independent blind vision-language model reads. Transcriptions were retained only when reads agreed after NFC normalization.
3. Visual review and audit: Visual inspection audited all real line crops across all splits. Label shift cascades caused by stray margin crops were realigned to printed page text. Leaked internal identifier placeholders were replaced with true text. Non-text ornamental dividers, noise specks, and sliced bounding boxes were assigned empty reference labels ("") with nonlinguistic category tags.
4. Deduplication and leakage prevention: All images were audited using global SHA256 image hashes and grayscale pixel values. Exact duplicate images were removed from training. Document pages are strictly isolated between training and evaluation splits so that no evaluation page appears in training data.

Schema and Columns

Column	Type	Description
id	string	Unique record identifier
image	struct	Image data dictionary containing bytes (PNG/JPEG) and path (null)
reference	string	Normalized text transcription
data_source	string	Origin of sample: "real" or "synthetic"
category	string	Script classification: "hindi", "english", "mixed_hindi_english", or "nonlinguistic"
document	string	Source book or document group identifier
dataset_version	string	Dataset release status: "final"
width	int32	Crop width in pixels
height	int32	Crop height in pixels

Usage

from datasets import Image, load_dataset

# Load validation split
val = load_dataset("hari31416/hindi-english-line-ocr", split="validation")
val = val.cast_column("image", Image())
print(val[0]["reference"])

# Load test split
test = load_dataset("hari31416/hindi-english-line-ocr", split="test")

# Load train split
train = load_dataset("hari31416/hindi-english-line-ocr", split="train")

Metadata Files

The repository provides sidecar metadata in the metadata/ directory:

• metadata/annotations.jsonl: Line bounding boxes, source URLs, page numbers, and transcription provenance.
• metadata/collection-summary.json: Document counts, candidate acceptance metrics, and layout breakdowns.
• metadata/label-corrections.json: Record of all reviewed label corrections, old labels, new labels, and reasons.
• metadata/leakage-audit.json: Deduplication audit and page isolation validation report.
• metadata/sources.json: Source publication URLs, SHA256 checksums, and document group mapping.
• metadata/characters.json: Inventory of observed characters.
• metadata/files.json: SHA256 checksums and file sizes for all dataset artifacts.
• vocabulary.json: Character vocabulary with 262 tokens for model training.