Hugging Face's logo

Datasets:
shantipriya
/
odia-ocr-merged

Modalities:
Image
Text
Formats:
parquet
Size:
10K - 100K
Libraries:
Datasets
Dask
Polars
Dataset card Data Studio Files
xet
 Community
1
odia-ocr-merged / README.md
shantipriya's picture
shantipriya
Update README.md
fb76ca2 verified
preview code
|
Raw History Blame Contribute Delete
8.19 kB
metadata 
dataset_info:
  features:
    - name: text
      dtype: string
    - name: image
      dtype: image
  splits:
    - name: train
      num_bytes: 13302255864
      num_examples: 58720
    - name: validation
      num_bytes: 1662781983
      num_examples: 7340
    - name: test
      num_bytes: 1662781983
      num_examples: 7340
  download_size: 16628681287
  dataset_size: 16627819830
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

Odia OCR - Merged Multi-Source Dataset

Overview

This is a comprehensive merged Odia Optical Character Recognition (OCR) dataset combining three major public datasets:

  1. OdiaGenAIOCR/Odia-lipi-ocr-data (64 samples)
  2. tell2jyoti/odia-handwritten-ocr (182,152 samples)
  3. darknight054/indic-mozhi-ocr - Odia subset (10,000+ samples)

Total: 192,000+ Odia OCR samples ready for training!

Dataset Contents

Source Breakdown

Dataset Samples Type License
OdiaGenAIOCR 64 Word-level documents Open Source
tell2jyoti 182,152 Character-level (32x32px) MIT
darknight054 10,000+ Printed word images Academic
TOTAL 192,000+ Mixed Open

Data Types

  1. Word-level OCR: Full page/document images with Odia text
  2. Character-level: Individual 32x32 grayscale Odia character images (47 OHCS characters)
  3. Printed Words: Professional printed Odia words from publications

Features

  • ✅ Mixed granularity: word-level, character-level, document-level
  • ✅ All 47 Odia characters represented
  • ✅ Balanced handwritten and printed text
  • ✅ Ready for immediate training

Loading the Dataset

From HuggingFace Hub (Recommended) 

from datasets import load_dataset

dataset = load_dataset("shantipriya/odia-ocr-merged")
train_data = dataset["train"]

From Local Directory 

from datasets import load_dataset

dataset = load_dataset("parquet", data_files="data.parquet")

Usage Examples

Basic Loading 

from datasets import load_dataset

dataset = load_dataset("shantipriya/odia-ocr-merged")
print(f"Total samples: {len(dataset['train'])}")

# Inspect first sample
first_sample = dataset['train'][0]
print(first_sample.keys())

PyTorch DataLoader 

from datasets import load_dataset
from torch.utils.data import DataLoader

dataset = load_dataset("shantipriya/odia-ocr-merged")
train_data = dataset['train']

def collate_fn(batch):
    images = [item['image'] for item in batch]
    texts = [item['text'] for item in batch]
    return {'images': images, 'texts': texts}

loader = DataLoader(train_data, batch_size=32, collate_fn=collate_fn)

for batch in loader:
    print(f"Batch: {len(batch['images'])} images")
    break

Data Splits 

from datasets import load_dataset
from sklearn.model_selection import train_test_split

dataset = load_dataset("shantipriya/odia-ocr-merged")
data = dataset['train']

# Create 80/10/10 split
train_size = int(0.8 * len(data))
val_size = int(0.1 * len(data))

train_data = data.select(range(train_size))
remaining = data.select(range(train_size, len(data)))

val_data = remaining.select(range(len(remaining) // 2))
test_data = remaining.select(range(len(remaining) // 2, len(remaining)))

print(f"Train: {len(train_data)}")
print(f"Val: {len(val_data)}")
print(f"Test: {len(test_data)}")

Training with Transformers

Fine-tuning Qwen2.5-VL 

from transformers import AutoProcessor, Qwen2VLForConditionalGeneration, TrainingArguments, Trainer
from datasets import load_dataset
from peft import LoraConfig, get_peft_model

# Load model
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct")
model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct")

# LoRA Configuration
lora_config = LoraConfig(
    r=32,
    lora_alpha=64,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
)
model = get_peft_model(model, lora_config)

# Load dataset
dataset = load_dataset("shantipriya/odia-ocr-merged")

# Training arguments
training_args = TrainingArguments(
    output_dir="./models/qwen-odia-ocr-v2",
    num_train_epochs=3,
    max_steps=500,
    warmup_steps=50,
    learning_rate=1e-4,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    save_steps=50,
    logging_steps=10,
    lr_scheduler_type="cosine",
    eval_strategy="steps",
    eval_steps=50,
)

# Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset['train'],
)

trainer.train()

Training Recommendations

Quick Test (1-2 hours) 

max_steps: 100
learning_rate: 5e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 10
scheduler: linear

Expected CER: 30-50%

Good Results (4-8 hours) 

max_steps: 500
learning_rate: 1e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 50
scheduler: cosine

Expected CER: 10-25%

Production Training (1-2 weeks) 

max_steps: 2000
learning_rate: 5e-5
batch_size: 2
gradient_accumulation_steps: 2
warmup_steps: 200
scheduler: cosine
eval_strategy: steps
save_steps: 100

Expected CER: 5-15%

Dataset Statistics

Sample Distribution

  • Word-level: 64 samples
  • Character-level: 182,152 samples
  • Printed words: 10,000+ samples

Character Coverage

Coverage of all 47 Odia characters from OHCS (Odia Handwritten Character Set):

  • Vowels: ଅ, ଆ, ଇ, ଈ, ଉ, ଊ, ଋ, ୠ, ଏ, ଐ, ଓ, ଔ (12)
  • Consonants: 33+ characters
  • Special marks: ୍, ଂ, ଃ

Citation

If you use this dataset, please cite the original sources:

@dataset{odia_ocr_merged_2026,
  title={Odia OCR - Merged Multi-Source Dataset},
  author={Parida, Shantipriya},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/shantipriya/odia-ocr-merged}}
}

@dataset{odiagenaiocr_2024,
  title={Odia-lipi-ocr-data},
  author={OdiaGenAIOCR},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/OdiaGenAIOCR/Odia-lipi-ocr-data}}
}

@inproceedings{gongidi2021iiit,
  title     = {IIIT-Indic-HW-Words: A Dataset for Indic Handwritten Text Recognition},
  author    = {Gongidi, Santhoshini and Jawahar, C. V.},
  booktitle = {International Conference on Document Analysis and Recognition (ICDAR)},
  pages     = {444--459},
  year      = {2021}
}

@inproceedings{dutta2018towards,
  title     = {Towards Spotting and Recognition of Handwritten Words in Indic Scripts},
  author    = {Dutta, Kartik and Krishnan, Praveen and Mathew, Minesh and Jawahar, C. V.},
  booktitle = {International Conference on Frontiers in Handwriting Recognition (ICFHR)},
  year      = {2018}
}

@dataset{odia_handwritten_ocr_2026,
  title={Odia Handwritten OCR Dataset},
  author={Jyoti},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/tell2jyoti/odia-handwritten-ocr}}
}

@dataset{darknight054_indic_mozhi_2024,
  title={Indic Mozhi OCR},
  author={darknight054},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/darknight054/indic-mozhi-ocr}}
}

License

This merged dataset combines:

  • OdiaGenAIOCR: Open Source
  • tell2jyoti: MIT License
  • darknight054: Academic License (per CVIT IIIT)

Please respect all individual licenses when using this dataset.

Contributors

  • Dataset Merging: Shantipriya Parida
  • Original Sources:
    • OdiaGenAIOCR team
    • tell2jyoti
    • CVIT IIIT (darknight054)

Contact

Related Resources

Last Updated: February 22, 2026 Version: 1.0.0 Status: Ready for Training