hheiden/Colored_Background_OCSR_benchmark

Dataset Card for Alpha-Extractor OCSR Benchmark

This dataset is a small, challenging benchmark for Optical Chemical Structure Recognition (OCSR), featuring 200 images. The data is a subset of the test data created for the "αExtractor" system, a tool specifically designed to perform well on difficult images, including those with backgrounds, low quality, and even hand-drawn structures. It serves as an excellent test of model robustness.

Dataset Details

Dataset Description

The Alpha-Extractor OCSR Benchmark is derived from the test data of the αExtractor system, a project by AlphaMa and the Shanghai Institute of Materia Medica (SIMM). The αExtractor recognition model is an enhanced version of a winning entry in the Bristol-Myers Squibb Molecular Translation Kaggle competition, designed for high accuracy on challenging, real-world chemical images.

The original test data collection was created to showcase the system's ability to handle images that conventional OCSR tools often struggle with. This Hugging Face dataset, prepared by Hunter Heidenreich, consists of a 200-image subset from that collection. It has been processed to include not only the source images and MOL files but also derived canonical SMILES, InChI, and SELFIES strings, making it ready for machine learning applications.

• Curated by: Original data curated by the alpha-Extractor team. This Hugging Face version was prepared by Hunter Heidenreich.
• License: MIT License.

Dataset Sources

• Repository:
  • Hugging Face Dataset Repo
  • Original Test Data GitHub Repository
• Paper:
  • αExtractor: a system for automatic extraction of chemical information from biomedical literature
• Demo:
  • αExtractor Online Server

Uses

Direct Use

Given its small size and challenging nature, this dataset is best suited for evaluating the robustness of pre-trained OCSR models. It is ideal for:

• Testing model performance on non-standard images (e.g., low-quality, hand-drawn).
• Error analysis to identify weaknesses in an OCSR pipeline.
• Benchmarking against a set of known difficult cases.

Out-of-Scope Use

This dataset is not suitable for training deep learning models from scratch due to its very small size (200 examples). Using it as a primary training set would lead to severe overfitting. Its specific and challenging composition may not be representative of any single, large-scale document source like patents or journals.

Dataset Structure

The dataset contains a single train split with 200 examples. Each example has the following fields:

• id (string): A unique identifier for the example.
• image (image): A PIL-encoded image of the chemical structure.
• mol (string): The ground truth structure in MOL file format.
• smiles (string): The canonical SMILES string for the molecule, generated from the mol data using RDKit.
• selfies (string): The SELFIES representation of the molecule, generated from the smiles string.
• inchi (string): The standard InChI string for the molecule, generated from the mol data using RDKit.

Dataset Creation

Curation Rationale

The original test data was curated to validate the performance of the αExtractor system, specifically its ability to generalize to diverse and low-quality chemical structure images where other systems might fail. This subset provides a standardized benchmark for these challenging cases.

Source Data

Data Collection and Processing

The source data was collected by the alpha-Extractor team and is available in their public GitHub repository. The repository contains a variety of image types, including standard benchmarks, images from the Bristol-Myers Squibb Kaggle competition, and specially collected examples of low-quality and hand-drawn structures.

This Hugging Face dataset is a 200-image subset of that larger collection. It was prepared by Hunter Heidenreich, who processed the ground truth MOL files using RDKit to generate the corresponding SMILES, InChI, and SELFIES strings.

Who are the source data producers?

The dataset was collected and curated by the research team behind the alpha-Extractor project.

Bias, Risks, and Limitations

• Extremely Small Size: With only 200 examples, any performance metrics calculated on this dataset should be interpreted with caution, as they may not be statistically significant.
• Purpose-Built Bias: The dataset was likely curated to contain a high proportion of "hard" examples to highlight the strengths of a particular model. It is therefore not a random or representative sample of chemical images.
• Unknown Composition: The exact breakdown of image types (e.g., how many are hand-drawn vs. low-quality) within this 200-image subset is not specified.

Recommendations

Use this dataset strictly for evaluation and robustness testing. It is a valuable tool for identifying failure modes in OCSR models but should not be used to draw broad conclusions about a model's overall performance without considering results from larger, more balanced benchmarks.

Citation

If you use this dataset, please cite the original αExtractor paper and this Hugging Face dataset to ensure reproducibility.

BibTeX:

@article{zhang2022extractor,
  title={αExtractor: a system for automatic extraction of chemical information from biomedical literature},
  author={Zhang, Rui and Chen, Zerun and Liu, Zaiyi and Zhang, Yanli and Wang, Xiaohong and Ai, Haitao and Liu, Jing and Du, Yong and Li, Youyong and Zhu, Weiliang},
  journal={Journal of Cheminformatics},
  volume={14},
  number={1},
  pages={64},
  year={2022},
  publisher={Springer}
}

@misc{huggingface_dataset_alphaextractor,
  author = {Heidenreich, Hunter},
  title = {Alpha-Extractor OCSR Benchmark},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{[https://huggingface.co/datasets/hheiden/alpha-extractor_OCSR_benchmark](https://huggingface.co/datasets/hheiden/alpha-extractor_OCSR_benchmark)}}
}