PatSnap/hiro-ocsr-real-24k

Hiro-OCSR Real 24K

Hiro-OCSR Real 24K is a real-world optical chemical structure recognition dataset for converting chemical structure images into machine-readable SMILES representations.

The dataset contains approximately 24,000 real chemical structure image samples with corresponding structure annotations. It is intended for training, evaluating, and benchmarking OCSR systems, especially models that need to handle chemical structure images from practical document scenarios.

Dataset Description

Optical Chemical Structure Recognition (OCSR) is the task of translating a chemical structure image into a structured molecular representation such as SMILES.

This dataset is released as part of the Hiro-OCSR project and focuses on real-world chemical structure images rather than synthetic-only renderings.

Intended Uses

This dataset can be used for:

• Training image-to-SMILES OCSR models
• Evaluating chemical structure recognition systems
• Benchmarking robustness on real-world chemistry images
• Research on chemical image understanding and document mining

Dataset Structure

Each sample contains a chemical structure image and its corresponding molecular structure annotation.

Typical fields include:

• image: chemical structure image
• SMILES or smiles: target molecular representation

Depending on the file format used in this repository, images may be stored directly in dataset records or referenced from image files.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("Patsnap/hiro-ocsr-real-24k")
print(dataset)

Task

The primary task is image-to-text generation:

chemical structure image -> SMILES

Limitations

The dataset is designed for OCSR research and evaluation. Model performance may vary depending on image quality, chemical notation style, abbreviations, R-groups, Markush-like structures, salts, stereochemistry, and other complex chemical drawing conventions. Users should validate generated SMILES with cheminformatics tooling such as RDKit when using predictions in downstream scientific workflows.

See DATASET DISCLAIMER.md for important limitations and responsible-use terms.

License

This dataset is released under the Apache License 2.0.

Citation

If you use this dataset, please cite the dataset repository:

@dataset{hiro_ocsr_real_24k,
  title = {Hiro-OCSR Real 24K},
  author = {Hiro-OCSR Contributors},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Patsnap/hiro-ocsr-real-24k}
}