OCR-Finetuning-EN-Dataset

A large-scale English OCR fine-tuning dataset containing synthetic and real-world text images for training modern OCR recognition models.

The dataset is distributed in Apache Parquet format with embedded image data, making it fully compatible with the Hugging Face datasets library and the Hugging Face Dataset Viewer.

---

Features

• ✅ 167,330 OCR image-text pairs
• ✅ Images embedded directly inside Parquet files
• ✅ Compatible with Hugging Face Dataset Viewer
• ✅ Works directly with datasets.load_dataset()
• ✅ Optimized for efficient downloading and streaming
• ✅ Ready for OCR model fine-tuning

---

Supported Models

This dataset is suitable for training and fine-tuning models including:

• PaddleOCR
• PARSeq
• TrOCR
• CRNN
• ABINet
• VisionEncoderDecoder Models
• Donut OCR
• Any sequence-based OCR recognition architecture

---

Dataset Statistics

Split	Samples
Train	150,597
Test	16,733
Total	167,330

Train/Test Split

• Train: 90%
• Test: 10%
• Random Seed: 42

---

Dataset Structure

OCR-Finetuning-EN-Dataset/

├── train/
│   └── data.parquet
│
├── test/
│   └── data.parquet
│
├── README.md
└── .gitattributes

---

Dataset Schema

Each row contains:

Column	Type	Description
id	int64	Sample identifier
image	Image	Embedded image
text	string	Ground-truth OCR transcription

Example:

{
    "id": 135714,
    "image": <PIL.Image>,
    "text": "3.00"
}

---

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "Srijan-Chakraborty/OCR-Finetuning-EN-Dataset"
)

print(dataset)

Access a sample:

sample = dataset["train"][0]

sample["image"].show()
print(sample["text"])

---

Exporting Back to Images + JSON

Although the dataset is distributed in Parquet format, it can easily be converted back into the traditional structure:

train/
│
├── images/
└── annotations.json

test/
│
├── images/
└── annotations.json

This allows easy integration with OCR frameworks that expect image folders and annotation JSON files.

Example:

import json
import os

from datasets import load_dataset

dataset = load_dataset(
    "Srijan-Chakraborty/OCR-Finetuning-EN-Dataset"
)

for split in dataset.keys():

    os.makedirs(f"{split}/images", exist_ok=True)

    annotations = []

    for sample in dataset[split]:

        filename = f"{sample['id']:06d}.jpg"

        sample["image"].save(
            os.path.join(split, "images", filename)
        )

        annotations.append(
            {
                "id": sample["id"],
                "file_name": filename,
                "text": sample["text"]
            }
        )

    with open(
        os.path.join(split, "annotations.json"),
        "w",
        encoding="utf-8"
    ) as f:
        json.dump(
            {
                "annotations": annotations
            },
            f,
            ensure_ascii=False,
            indent=4
        )

---

Intended Use

This dataset is intended for:

• OCR Recognition
• OCR Fine-tuning
• OCR Benchmarking
• Scene Text Recognition
• Document OCR
• Vision-Language Research
• Sequence Recognition
• OCR Pretraining
• OCR Evaluation

---

Citation

If you use this dataset in your research, please cite:

@misc{ocr_finetuning_en_dataset,
  author = {Srijan Chakraborty},
  title = {OCR-Finetuning-EN-Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Srijan-Chakraborty/OCR-Finetuning-EN-Dataset}
}

---

Author

Srijan Chakraborty

GitHub: https://github.com/SrijanChakraborty2003

Hugging Face: https://huggingface.co/Srijan-Chakraborty

---

License

This dataset is released under the Apache License 2.0.

The dataset was constructed by combining publicly available OCR datasets. Please ensure that the licenses and usage terms of the original source datasets are respected when using this dataset.