Phitran21/synthetic-ocr-en-det-rec-120k

Synthetic English OCR Detection and Recognition 240K

## 📌 Current dataset size: 240,000 paired OCR samples The current v2.0 release contains exactly 240,000 detector images and 240,000 matching recognition crops. Each sample ID corresponds to: - one full image for text detection; - one cropped text image for text recognition; - one detector JSONL record; - one recognizer JSONL record. Therefore, the dataset contains 240,000 aligned OCR pairs and 480,000 JPEG files in total. The detector image and recognition crop are two representations of the same OCR sample, not 480,000 independent texts. The repository identifier still ends in synthetic-ocr-en-det-rec-120k because that was the name of the original release. The old repository name is intentionally retained to preserve existing links, citations, download scripts, and bookmarks. The repository name is legacy; the current dataset size is 240K.

A large synthetic English OCR dataset containing 240,000 full detector images, 240,000 cropped text images, polygon-based text annotations, and aligned transcription labels.

This dataset is designed for training and evaluating:

• —OCR text detection
• —OCR text recognition
• —End-to-end OCR systems
• —Scene text recognition
• —Document text extraction
• —Polygon-based text localization
• —Lightweight mobile and ONNX OCR models

The dataset includes two aligned subsets:

1. 1.Detector dataset for locating text regions in full images.
2. 2.Recognizer dataset for converting cropped text images into English text.

Both subsets use the same numeric sample IDs. For example, images/en_00000001.jpg and crops/en_00000001.jpg belong to the same OCR sample and share the same ground-truth transcription.

Dataset Summary

The images contain English text rendered using different fonts, positions, sizes, rotations, backgrounds, colors, and visual styles.

The goal is to provide a practical OCR training resource for researchers and developers building text detection and recognition systems.

Repository Name Notice

The Hugging Face repository URL remains:

Phitran21/synthetic-ocr-en-det-rec-120k

The 120k suffix refers to the original release size. The dataset was expanded in place to 240,000 paired samples so that existing users would not lose access through old links or citations.

Use the following values when describing the current release:

Current title: Synthetic English OCR Detection and Recognition 240K
Current version: v2.0
Current paired samples: 240,000
Legacy repository slug: synthetic-ocr-en-det-rec-120k

Repository Structure

synthetic-ocr-en-det-rec-120k/
├── detector/
│   ├── detector.jsonl
│   └── images/
│       ├── part_001.zip
│       ├── part_002.zip
│       ├── part_003.zip
│       ├── part_004.zip
│       ├── part_005.zip
│       ├── part_006.zip
│       ├── part_007.zip
│       ├── part_008.zip
│       ├── part_009.zip
│       └── part_010.zip
│
└── recognizer/
    ├── recognizer.jsonl
    └── crops/
        ├── part_001.zip
        ├── part_002.zip
        ├── part_003.zip
        ├── part_004.zip
        ├── part_005.zip
        ├── part_006.zip
        ├── part_007.zip
        ├── part_008.zip
        ├── part_009.zip
        └── part_010.zip

The ZIP archives are divided into approximately equal parts to make downloading, storage, verification, and extraction easier.

The v2.0 extension is stored in part_006.zip through part_010.zip:

File Descriptions

detector/detector.jsonl

This file contains annotations for training a text detection model.

Each JSON line represents one full image and includes:

• —Image path
• —Image width and height
• —Text polygon coordinates
• —Ground-truth transcription
• —Language
• —Text direction
• —Confidence
• —Background information
• —Font information
• —Synthetic-data indicator

Example:

{
  "image": "images/en_00000001.jpg",
  "width": 960,
  "height": 640,
  "items": [
    {
      "polygon": [
        [676, 144],
        [682, 146],
        [688, 149]
      ],
      "text": "Example English text.",
      "lang": "en",
      "direction": "horizontal",
      "confidence": 1.0
    }
  ],
  "synthetic": true
}

The complete polygon contains multiple coordinate points that describe the text boundary more precisely than a simple rectangular bounding box.

Use this file together with:

detector/images/part_*.zip

After extraction, the image paths have the following structure:

images/en_00000001.jpg
images/en_00000002.jpg
...
images/en_00240000.jpg

recognizer/recognizer.jsonl

This file contains annotations for training a text recognition model.

Each JSON line represents one cropped text image and includes:

• —Crop image path
• —Ground-truth text
• —Language
• —Text direction
• —Original source image
• —Original source polygon
• —Crop box
• —Rotation angle

Example:

{
  "image": "crops/en_00000001.jpg",
  "text": "Example English text.",
  "lang": "en",
  "direction": "horizontal",
  "source_image": "images/en_00000001.jpg",
  "source_polygon": [
    [676, 144],
    [682, 146],
    [688, 149]
  ],
  "crop_box": [39, 92, 767, 314],
  "rotation": -3.257
}

Use this file together with:

recognizer/crops/part_*.zip

After extraction, the crop paths have the following structure:

crops/en_00000001.jpg
crops/en_00000002.jpg
...
crops/en_00240000.jpg

Pair Alignment

Detector and recognizer records are aligned by their numeric IDs:

Detector image:     images/en_00012345.jpg
Recognizer crop:   crops/en_00012345.jpg
Detector record:   detector JSONL record for ID 00012345
Recognizer record: recognizer JSONL record for ID 00012345
Ground-truth text: identical in both records

This alignment allows the two subsets to be trained independently or combined in an end-to-end OCR pipeline.

Detection Dataset Format

The detector subset follows this relationship:

Full image
    ↓
Polygon annotation
    ↓
Text region localization

Training input:

images/en_XXXXXXXX.jpg

Training target:

items[].polygon

Optional transcription information is available in:

items[].text

The detector annotations are suitable for:

• —DBNet
• —Differentiable Binarization OCR
• —CRAFT-style detectors
• —EAST-style detectors
• —Segmentation-based OCR detection
• —Polygon regression models
• —Custom object detection pipelines

Some frameworks require four-point quadrilaterals or rectangular boxes. In that case, polygon coordinates can be converted during preprocessing.

Recognition Dataset Format

The recognizer subset follows this relationship:

Cropped text image
    ↓
OCR recognition model
    ↓
English transcription

Training input:

crops/en_XXXXXXXX.jpg

Training target:

text

This subset can be used with:

• —CRNN
• —CTC-based OCR
• —Transformer OCR
• —Attention-based recognition
• —SVTR
• —PARSeq-style systems
• —PaddleOCR recognition models
• —ONNX Runtime OCR pipelines
• —Mobile OCR applications

Downloading the Dataset

Using Git

git lfs install

git clone \
  https://huggingface.co/datasets/Phitran21/synthetic-ocr-en-det-rec-120k

Using the Hugging Face CLI

pip install -U huggingface_hub

hf download \
  Phitran21/synthetic-ocr-en-det-rec-120k \
  --repo-type dataset \
  --local-dir synthetic-ocr-en-det-rec-240k

The local directory may use 240k even though the stable remote repository slug still uses 120k.

Extracting the Detector Images

Linux or Ubuntu:

mkdir -p extracted_detector

for file in detector/images/part_*.zip; do
  unzip "$file" -d extracted_detector
done

Result:

extracted_detector/
└── images/
    ├── en_00000001.jpg
    ├── en_00000002.jpg
    ├── ...
    └── en_00240000.jpg

The corresponding annotation file is:

detector/detector.jsonl

Extracting the Recognition Crops

mkdir -p extracted_recognizer

for file in recognizer/crops/part_*.zip; do
  unzip "$file" -d extracted_recognizer
done

Result:

extracted_recognizer/
└── crops/
    ├── en_00000001.jpg
    ├── en_00000002.jpg
    ├── ...
    └── en_00240000.jpg

The corresponding annotation file is:

recognizer/recognizer.jsonl

Quick JSONL Inspection

Inspect the first detector samples:

head -n 5 detector/detector.jsonl

Inspect the first recognition samples:

head -n 5 recognizer/recognizer.jsonl

Validate JSONL using Python:

import json
from pathlib import Path


def validate_jsonl(path: str) -> int:
    count = 0

    with Path(path).open("r", encoding="utf-8") as file:
        for line_number, line in enumerate(file, start=1):
            try:
                json.loads(line)
                count += 1
            except json.JSONDecodeError as error:
                raise ValueError(
                    f"Invalid JSON at line {line_number}: {error}"
                ) from error

    return count


print(
    "Detector records:",
    validate_jsonl("detector/detector.jsonl"),
)

print(
    "Recognizer records:",
    validate_jsonl("recognizer/recognizer.jsonl"),
)

Expected result for v2.0:

Detector records: 240000
Recognizer records: 240000

Python Loading Example

Load recognition annotations:

import json
from pathlib import Path


annotation_path = Path("recognizer/recognizer.jsonl")
samples = []

with annotation_path.open("r", encoding="utf-8") as file:
    for line in file:
        samples.append(json.loads(line))

print("Number of samples:", len(samples))
print("First sample:", samples[0])

Load detector annotations:

import json
from pathlib import Path


annotation_path = Path("detector/detector.jsonl")

with annotation_path.open("r", encoding="utf-8") as file:
    first_sample = json.loads(next(file))

print("Image:", first_sample["image"])
print("Width:", first_sample["width"])
print("Height:", first_sample["height"])
print("Text items:", len(first_sample["items"]))
print("First polygon:", first_sample["items"][0]["polygon"])

Suggested Train, Validation, and Test Split

The dataset is currently distributed as one complete collection.

A recommended split is:

Split by filename or record index using a fixed random seed to ensure reproducibility.

Example:

import json
import random
from pathlib import Path


random.seed(42)
source = Path("recognizer/recognizer.jsonl")

with source.open("r", encoding="utf-8") as file:
    rows = [json.loads(line) for line in file]

random.shuffle(rows)

total = len(rows)
train_end = int(total * 0.90)
validation_end = int(total * 0.95)

splits = {
    "train": rows[:train_end],
    "validation": rows[train_end:validation_end],
    "test": rows[validation_end:],
}

for split_name, split_rows in splits.items():
    output = Path(f"recognizer_{split_name}.jsonl")

    with output.open("w", encoding="utf-8") as file:
        for row in split_rows:
            file.write(
                json.dumps(row, ensure_ascii=False) + "\n"
            )

    print(split_name, len(split_rows))

When creating detector and recognizer splits, use the same sample IDs for both subsets so that pair alignment is preserved.

Intended Uses

This dataset is intended for:

• —Academic OCR research
• —Non-commercial OCR model training
• —OCR benchmarking
• —Text detection experiments
• —Text recognition experiments
• —Synthetic-data research
• —Document AI research
• —Mobile OCR development
• —ONNX and ONNX Runtime experiments
• —Educational projects
• —Personal non-commercial projects

Out-of-Scope Uses

The dataset must not be used for:

• —Commercial use without written permission
• —Illegal surveillance
• —Privacy-invasive identification systems
• —Misleading or fraudulent applications
• —Applications that violate applicable laws
• —Claiming the dataset was manually collected or manually annotated
• —Redistributing the dataset under incompatible terms

Limitations

This is a synthetic dataset and does not fully represent all real-world OCR conditions.

Possible limitations include:

• —Synthetic fonts and rendering patterns
• —Limited background diversity
• —Limited handwriting coverage
• —Limited severe blur and compression artifacts
• —Limited curved or highly distorted text
• —Possible unnatural source sentences
• —Possible differences from photographs taken by real cameras
• —Possible imbalance among fonts, rotations, text lengths, and backgrounds
• —Primarily horizontal English text
• —No guarantee of perfect semantic or grammatical quality in every sentence

Models trained exclusively on this dataset may require fine-tuning on real-world OCR data before production use.

For stronger generalization, consider combining this dataset with legally compatible real-image datasets.

Data Quality Notes

The annotations were generated automatically as part of the synthetic rendering process.

Because the text, polygon, crop, and transcription originate from the same generation pipeline, labels are expected to align closely with their corresponding images.

The v2.0 release was packaged with:

• —Aligned detector and recognizer IDs
• —Matching detector and recognizer text
• —Continuous IDs from 00000001 through 00240000
• —JSONL syntax validation
• —Missing-file checks
• —Duplicate-ID checks
• —Duplicate-text checks for the newly generated extension
• —ZIP path and CRC verification

Users should still perform their own validation before training production systems.

Recommended checks include:

• —ZIP integrity
• —Missing file detection
• —Duplicate file detection
• —JSONL parsing
• —Image readability
• —Polygon coordinate bounds
• —Empty transcription detection
• —Train and test leakage detection

License

This dataset is licensed under the:

Creative Commons Attribution-NonCommercial 4.0 International License CC BY-NC 4.0

You may:

• —Use the dataset for research
• —Use the dataset for education
• —Use the dataset for personal projects
• —Modify and adapt the dataset
• —Train non-commercial models
• —Redistribute permitted adaptations with attribution

You must:

• —Credit the original creator
• —Link or refer to this dataset repository
• —Clearly indicate significant modifications
• —Keep attribution information visible
• —Comply with the CC BY-NC 4.0 license

You may not:

• —Use the dataset commercially without prior written permission
• —Sell the dataset or access to the dataset
• —Include the dataset in a paid commercial product without permission
• —Use the dataset to provide a paid OCR service without permission
• —Re-license the original dataset under incompatible terms
• —Claim ownership of the original dataset

Commercial Licensing

Commercial use is not included under the public CC BY-NC 4.0 license.

For commercial use, enterprise use, paid products, paid APIs, paid applications, commercial model training, or commercial redistribution, prior written permission is required.

Please contact the creator to discuss a separate commercial license.

Attribution

Suggested citation:

Synthetic English OCR Detection and Recognition 240K
Version 2.0
Created by Trần Phi
Hugging Face: Phitran21/synthetic-ocr-en-det-rec-120k
License: CC BY-NC 4.0

Suggested attribution for model cards:

This model was trained using the Synthetic English OCR Detection
and Recognition 240K dataset (v2.0), created by Trần Phi:
https://huggingface.co/datasets/Phitran21/synthetic-ocr-en-det-rec-120k

The repository URL contains the legacy 120k identifier, while the current release contains 240,000 paired samples.

Creator

Trần Phi

Hugging Face:

https://huggingface.co/Phitran21

Contact

For dataset questions, issue reports, collaboration, or commercial licensing:

• —Email: phihhhhhhhhhh@gmail.com
• —Facebook: https://www.facebook.com/share/1PqDzPQJYf/
• —Website: https://toren.io.vn

Email responses may be limited or delayed. For public technical questions, using the Hugging Face Community tab is recommended.

Reporting Issues

When reporting a problem, please include:

• —Affected file name
• —ZIP part name
• —JSONL line number
• —Description of the issue
• —Minimal reproduction steps
• —Screenshot or sample when appropriate

Please use the Community tab of this repository for public bug reports and technical discussions.

Version History

v2.0 — Current

• —240,000 detector images
• —240,000 recognition crops
• —240,000 aligned detector/recognizer pairs
• —480,000 JPEG files in total
• —Detector JSONL annotations
• —Recognition JSONL annotations
• —Ten detector ZIP archives
• —Ten recognizer ZIP archives
• —Continuous sample IDs through en_00240000.jpg

v1.0 — Legacy 120K release

• —Initial approximately 120K paired OCR release
• —Five detector ZIP archives
• —Five recognizer ZIP archives
• —Original repository name established as synthetic-ocr-en-det-rec-120k

The repository name was retained for backward compatibility when v2.0 expanded the dataset to 240K.

Acknowledgements

Thank you to the open-source OCR, computer vision, Python, font, and machine-learning communities whose tools and research make synthetic dataset creation possible.

If this dataset is useful in your research or project, please consider giving the repository a like and citing the dataset.