Faizaniqbal/IndicOCR

A large-scale, high-fidelity synthetic Document AI & OCR benchmark spanning 23 Pan-Indic languages and 12 distinct writing systems.

---

1. Overview

IndicOCR (IndicPixel) is a large-scale, multilingual Optical Character Recognition (OCR) and Document AI benchmark purpose-built for the South Asian linguistic ecosystem. Spanning all 22 Eighth Schedule Constitutional Languages of India plus Bhojpuri, the dataset covers 12 distinct writing systems including Devanagari, Bengali-Assamese, Gurmukhi, Gujarati, Odia, Tamil, Telugu, Kannada, Malayalam, Extended Perso-Arabic (RTL), Ol Chiki, and Meetei Mayek.

Why IndicOCR?

Traditional Indic OCR datasets frequently suffer from:

1. Diacritic Truncation: Top/bottom vowel signs (matras, halants, bindis) being clipped at canvas boundaries.
2. Improper Complex Text Layout (CTL): Broken conjuncts (samyuktaksharas) and half-forms from naive fallback rasterizers.
3. Sterile Renders: Synthetic datasets lacking authentic degradation physics (photostat artifacts, ink bleed, non-uniform paper textures, shadows).

IndicOCR resolves these foundational challenges by combining native HarfBuzz OpenType shaping, font-level vertical metric padding, and an Indian document degradation DAG with 54 physical print, substrate, and sensor operators.

---

2. Dataset Specifications

• Verified Volume: 13,250,000 paired samples packaged into 2,650 standardized WebDataset shards (.tar), with 5,000 samples per shard.
• Complex Text Layout (CTL) Guarantee: Shaped with uharfbuzz with explicit script and direction (ltr / rtl), mathematically enforcing zero Glyph ID 0 (.notdef tofu boxes).
• Diacritic Ownership Gate: Dynamic vertical zone calculation ensures 100% of ascenders, body zones, and descenders lie strictly within canvas boundaries.
• 4-Tier Structural Granularity Matrix:
  • Tier 1: Full-Page Document Spreads (2%) — Multi-column administrative gazettes, broadsheet newspapers, and literary journals with word-level bounding boxes.
  • Tier 2: Multi-Line Paragraph Blocks (13%) — Wrapped multi-line natural reading sequences.
  • Tier 3: Reading Lines (55%) — Core continuous text sentences with token-level bounding coordinates.
  • Tier 4: Isolated Tokens & Complex Conjuncts (30%) — Rare conjunct dictionaries, ligatures, numerals, and single words.
• Master 54-Operator Degradation Engine: 12.5% clean digital baseline scans, 87.5% realistic physical degradation across 5 distinct categories (paper substrates, print voiding, ink bleed, ancient manuscript wear, optical/sensor transformations).

---

3. Language Portfolio & Direct Hugging Face Partitions

Explore or stream individual language partitions directly on Hugging Face:

Language	Script	ISO Code	Verified Samples	POSIX Shards	Hugging Face Partition
Hindi	Devanagari (Deva)	hin	1,000,000	200	Explore data/hindi/ ↗
Urdu	Extended Arabic RTL (Arab)	urd	1,000,000	200	Explore data/urdu/ ↗
Bengali	Bengali (Beng)	ben	1,000,000	200	Explore data/bengali/ ↗
Tamil	Tamil (Taml)	tam	1,000,000	200	Explore data/tamil/ ↗
Marathi	Devanagari (Deva)	mar	1,000,000	200	Explore data/marathi/ ↗
Telugu	Telugu (Telu)	tel	500,000	100	Explore data/telugu/ ↗
Gujarati	Gujarati (Gujr)	guj	500,000	100	Explore data/gujarati/ ↗
Kannada	Kannada (Knda)	kan	500,000	100	Explore data/kannada/ ↗
Malayalam	Malayalam (Mlym)	mal	500,000	100	Explore data/malayalam/ ↗
Odia	Odia (Orya)	ori	500,000	100	Explore data/odia/ ↗
Punjabi	Gurmukhi (Guru)	pan	500,000	100	Explore data/punjabi/ ↗
Assamese	Bengali-Assamese (Beng)	asm	500,000	100	Explore data/assamese/ ↗
Nepali	Devanagari (Deva)	nep	500,000	100	Explore data/nepali/ ↗
Sanskrit	Devanagari (Deva)	san	500,000	100	Explore data/sanskrit/ ↗
Santali	Ol Chiki (Olck)	sat	500,000	100	Explore data/santali/ ↗
Manipuri	Meetei Mayek (Mtei)	mni	500,000	100	Explore data/manipuri/ ↗
Sindhi	Extended Arabic RTL (Arab)	snd	500,000	100	Explore data/sindhi/ ↗
Bodo	Devanagari (Deva)	brx	500,000	100	Explore data/bodo/ ↗
Bhojpuri	Devanagari (Deva)	bho	500,000	100	Explore data/bhojpuri/ ↗
Kashmiri	Extended Arabic RTL (Arab)	kas	500,000	100	Explore data/kashmiri/ ↗
Konkani	Devanagari (Deva)	gom	250,000	50	Explore data/konkani/ ↗
Maithili	Devanagari (Deva)	mai	250,000	50	Explore data/maithili/ ↗
Dogri	Devanagari (Deva)	doi	250,000	50	Explore data/dogri/ ↗
TOTAL	12 Writing Systems	23 Langs	13,250,000	2,650	Browse All Partitions on Hugging Face ↗

---

4. Structural Specification & Annotation Format

Each sample is packaged inside standardized WebDataset .tar archives containing paired lossless image files (.webp) and comprehensive metadata descriptors (.json):

{
  "sample_key": "kan_0008644",
  "tier": "Tier_2_Paragraph",
  "lang": "kan",
  "script": "Knda",
  "font_name": "BalooTamma2-Regular.ttf",
  "is_clean": false,
  "canvas_size": [669, 213],
  "text": "ಅದನ್ನು ತಂದು ಶ್ರೀನಿವಾಸನಿಗೆ ಕೊಡುತ್ತಾಳೆ. ಅಂಗಡಿಗೆ ಹಿಂದಿರುಗಿ\nಬಂದ ಶ್ರೀನಿವಾಸ ಡಬ್ಬಿಯನ್ನು",
  "tokens": [
    {"text": "ಅದನ್ನು", "bbox": [24, 16, 110, 68]},
    {"text": "ತಂದು", "bbox": [122, 16, 185, 68]},
    {"text": "ಶ್ರೀನಿವಾಸನಿಗೆ", "bbox": [198, 16, 360, 68]}
  ],
  "token_count": 22,
  "applied_augmentations": [
    "#04 Substrate: notebook_ruled",
    "#40 Keystone Tilt",
    "#13 Capillary Ink Bleed"
  ]
}

Metadata Fields

Field Name	Type	Description
sample_key	string	Unique globally indexed sample identifier across shards.
tier	string	Granularity tier (Tier_1_FullPage, Tier_2_Paragraph, Tier_3_Line, Tier_4_Word).
lang	string	ISO 639-3 three-letter language identifier.
script	string	ISO 15924 four-letter script identifier.
font_name	string	OpenType font family utilized for HarfBuzz shaping.
canvas_size	[int, int]	Canvas dimensions [width, height] in pixels.
text	string	Unicode NFC ground truth text transcription.
tokens	list[dict]	Token-level annotations with tight ink bounding boxes [x0, y0, x1, y1].
token_count	int	Total count of recognized tokens/words in the sample.
applied_augmentations	list[string]	Chain of degradation operators applied from the 54-operator DAG.

---

5. Distributed Streaming Quickstart (PyTorch / WebDataset)

IndicOCR shards stream directly into GPU memory over HTTP without local disk extraction:

import webdataset as wds
from torch.utils.data import DataLoader

# Stream Kannada shards directly from Hugging Face Hub
shard_url = "https://huggingface.co/datasets/Faizaniqbal/IndicOCR/resolve/main/data/kannada/kan_train_{00000..00099}.tar"

dataset = (
    wds.WebDataset(shard_url, resampled=False, shardshuffle=True)
    .shuffle(1000)
    .decode("pil")
    .to_tuple("webp", "json")
)

dataloader = DataLoader(dataset, batch_size=32, num_workers=4)

for images, metadata in dataloader:
    # Forward pass to Vision-Language Model / TrOCR / Donut / LayoutLMv3
    pass

---

6. License

The IndicOCR dataset and benchmark documentation are released under the Apache 2.0 License.

---

7. Citation

If you utilize IndicOCR in your research or applications, please cite:

@dataset{indicpixel_2026,
  author    = {Faizan Iqbal},
  title     = {IndicOCR: A Large-Scale Foundational Multilingual Document AI and OCR Benchmark for Pan-Indic Languages},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Faizaniqbal/IndicOCR}
}