RxnCaption-15K: Chemical Reaction Images and Annotations from Scientific Literature

15,400 chemical reaction images were extracted from scientific literature PDFs and annotated for diagram parsing. The dataset, created by songjhPKU and last updated in February 2026, is split into 15,000 training and 400 test images. It is designed to support research in cheminformatics, document analysis, and computer vision.

Use Cases

• Train image parsing models based on annotated chemical reaction diagrams.
• Develop document analysis tools for scientific PDFs based on extracted reaction images.
• Build multimodal models for chemistry based on the combination of images and detailed annotations.
• Benchmark computer vision algorithms for specialized diagram recognition tasks.

Strengths

• 15,400 total images provides a substantial corpus for model training.
• 15,000 images are designated for training, offering a large supervised learning set.
• Annotations are described as detailed, which may support complex parsing tasks.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The source and selection criteria for the scientific literature are not specified, which may introduce bias.
• The license is unknown, which may restrict commercial or certain research uses.

Provenance

Source

songjhPKU on Hugging Face.

Collection Method

Images extracted from scientific literature PDFs.

Time Range

null

Freshness

Last updated 2026-02-10 03:30:53.

Geography

null