Datasets:

Maksonchek
/

codeocr-dataset

dataset_info:
  features:
    - name: id
      dtype: string
    - name: difficulty
      dtype: string
    - name: code
      dtype: string
    - name: render_light
      dtype: image
    - name: render_dark
      dtype: image
    - name: photo
      dtype: image
  splits:
    - name: easy
      num_bytes: 3082543834
      num_examples: 700
    - name: medium
      num_bytes: 902595780
      num_examples: 200
    - name: hard
      num_bytes: 500128560
      num_examples: 100
  download_size: 4481644051
  dataset_size: 4485268174
configs:
  - config_name: default
    data_files:
      - split: easy
        path: data/easy-*
      - split: medium
        path: data/medium-*
      - split: hard
        path: data/hard-*
license: mit
task_categories:
  - image-to-text
  - text-generation
tags:
  - code
  - ocr
pretty_name: CodeOCR
size_categories:
  - 1K<n<10K

pretty_name: "CodeOCR Dataset (Python Code Images + Ground Truth)" license: mit language: - en task_categories: - image-to-text tags: - ocr - code - python - leetcode - synthetic - computer-vision size_categories: - 1K<n<10K

CodeOCR Dataset (Python Code Images + Ground Truth)

This dataset is designed for Optical Character Recognition (OCR) of source code.
Each example pairs Python code (ground-truth text) with image renderings of that code (light/dark themes) and a real photo.

Dataset Summary

Language: Python (text ground truth), images of code
Splits: easy, medium, hard
Total examples: 1,000
- easy: 700
- medium: 200
- hard: 100
Modalities: image + text

What is “ground truth” here?

The code field is exactly the content of gt.py used to generate the synthetic renderings.
During dataset creation, code is normalized to ensure stable GT properties:

UTF-8 encoding
newline normalization to LF (\n)
tabs expanded to 4 spaces
syntax checked with Python compile() (syntax/indentation correctness)

This makes the dataset suitable for training/evaluating OCR models that output plain code text.

Data Fields

Each row contains:

id (string): sample identifier (e.g., easy_000123)
difficulty (string): easy / medium / hard
code (string): ground-truth Python code
render_light (image): synthetic rendering (light theme)
render_dark (image): synthetic rendering (dark theme)
photo (image): real photo of the code

How to Use

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("maksonchek/codeocr-dataset")
print(ds)
print(ds["easy"][0].keys())

Access code and images

ex = ds["easy"][0]

# Ground-truth code
print(ex["code"][:500])

# Images are stored as `datasets.Image` features.
render = ex["render_light"]
print(render)

If your environment returns image dicts with local paths:

from PIL import Image

img = Image.open(ex["render_light"]["path"])
img.show()

Real photo (always present in this dataset):

from PIL import Image

photo = Image.open(ex["photo"]["path"])
photo.show()

Dataset Creation

1) Code selection

Python solutions were collected from an open-source repository of LeetCode solutions (MIT licensed).

2) Normalization to produce stable GT

The collected code is written into gt.py after:

newline normalization to LF
tab expansion to 4 spaces
basic cleanup (no hidden control characters)
Python syntax check via compile()

3) Synthetic rendering

Synthetic images are generated from the normalized gt.py in:

light theme (render_light)
dark theme (render_dark)

4) Real photos

Real photos are manually captured and linked for every sample.

Statistics (high-level)

Average code length by difficulty (computed on this dataset):

easy: ~27 lines, ~669 chars
medium: ~36 lines, ~997 chars
hard: ~55 lines, ~1767 chars

(Exact values may vary if the dataset is extended.)

Intended Use

OCR for programming code
robust text extraction from screenshot-like renders and real photos
benchmarking OCR pipelines for code formatting / indentation preservation

Not Intended Use

generating or re-distributing problem statements
competitive programming / cheating use-cases

Limitations

Code is checked for syntax correctness, but not necessarily for runtime correctness.
Rendering style is controlled and may differ from real-world photos.

License & Attribution

This dataset is released under the MIT License.

The included solution code is derived from kamyu104/LeetCode-Solutions (MIT License):
https://github.com/kamyu104/LeetCode-Solutions

If you use this dataset in academic work, please cite the dataset and credit the original solution repository.

Citation

BibTeX

@dataset{codeocr_leetcode_2025,
  author       = {Maksonchek},
  title        = {CodeOCR Dataset (Python Code Images + Ground Truth)},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/maksonchek/codeocr-dataset}
}