Patent Markush Structure Images And CXSMILES Representations

Four benchmark datasets contain images of chemical Markush structures from patents and their corresponding CXSMILES string representations. The largest subset, 'uspto-mol-m-54k-new', includes 54,785 training samples. The datasets were created by docling-project and were last updated in March 2026.

Use Cases

• Train image-to-text models to convert Markush structure images into CXSMILES strings using the 'uspto-mol-m-54k-new' training set.
• Evaluate model performance on the 'uspto-markush' benchmark of 74 ground truth OCR samples.
• Benchmark conversion accuracy on the 'm2s' subset of 103 Mol2Smiles (M2S) samples.
• Test model generalization on the 'IP5-markush' collection of 878 Markush structures from international patent offices.

Strengths

• Largest training subset contains 54,785 samples for model training.
• Provides four distinct benchmark subsets for evaluation, totaling over 1,200 test/benchmark samples.
• Includes ground truth OCR annotations for precise model evaluation.

Limitations

• Total dataset size and file formats are unspecified.
• Limited geographic and temporal coverage details are provided for the patent sources.
• The 'uspto-markush' and 'm2s' benchmark sets are relatively small (74 and 103 samples).

Provenance

Source

Primarily derived from USPTO (United States Patent and Trademark Office) and IP5 patent office documents.

Collection Method

Images of Markush structures extracted from patents, with corresponding CXSMILES representations generated.

Freshness

Last updated March 2026.

Geography

International, with focus on USPTO and IP5 patent offices.