Patent Markush Structure Images with CXSMILES Representations

Datasets contain images of Markush chemical structures from patents paired with their CXSMILES string representations. The collection includes over 54,000 training samples from the USPTO-MOL-M source and multiple benchmark subsets for evaluation. The dataset was created by docling-project and was last updated in March 2026.

Use Cases

• Train a model to convert patent Markush structure images into CXSMILES strings using the 54,785 training samples.
• Benchmark model performance on the 74 USPTO Markush structures using ground truth OCR labels.
• Evaluate model generalization on the 878 IP5 Markush structures from the benchmark subset.
• Compare model outputs against ChemicalOCR predictions provided for the USPTO-MOL-M subset.
• Test model accuracy on the 103 Mol2Smiles (M2S) benchmark samples.

Strengths

• Contains 54,785 training samples in the uspto-mol-m-54k-new subset.
• Includes multiple benchmark subsets (74, 103, and 878 samples) for structured evaluation.
• Provides both ground truth and predicted OCR labels for comparison.

Limitations

• The total number of rows across all subsets is not specified.
• The specific image formats, resolutions, and annotation quality are unknown.
• The dataset is focused on patent-derived structures, which may not represent other chemical domains.

Provenance

Source

Primarily derived from USPTO-MOL-M and other patent sources (USPTO, IP5).

Collection Method

Images of Markush structures were extracted from patents and paired with CXSMILES representations, with some subsets using ChemicalOCR predictions.

Freshness

Last updated March 2026.