docling-project/MarkushGrapher-2-Datasets

MarkushGrapher 2 Datasets

Datasets for training and evaluating MarkushGrapher 2, a model for converting patent Markush structure images into CXSMILES representations.

Dataset Subsets

Subset	Train	Test	Description	OCR
uspto-mol-m-54k-new	54,785	200	USPTO-MOL-M Markush samples	ChemicalOCR predictions
uspto-markush	—	74	USPTO Markush structures benchmark	Ground Truth OCR
m2s	—	103	Mol2Smiles (M2S) benchmark	Ground Truth OCR
IP5-markush	—	878	IP5 Markush structures benchmark	Ground Truth OCR

Features

Each sample contains:

• page_image — Input patent image (PIL Image, typically 1024×1024)
• cells — OCR-detected text cells with bounding boxes (bbox in normalized coordinates, text)
• cxsmiles — Ground truth CXSMILES representation
• cxsmiles_opt — Optimized (tokenizer-friendly) CXSMILES representation
• cxsmiles_dataset — Original CXSMILES from the source dataset
• annotation — Annotation metadata (used to train model)
• image_name — Source image filename
• id — Sample identifier

Usage

from datasets import load_dataset

# Load a specific subset
dataset = load_dataset("docling-project/MarkushGrapher-2-Datasets", "uspto-mol-m-54k")

# Load a benchmark subset
benchmark = load_dataset("docling-project/MarkushGrapher-2-Datasets", "m2s")

Note

MarkushGrapher-2 is also trained on the following datasets:

• Phase 1: 243k real-world image–SMILES pairs from MolScribe

• Phase 2:

  • 235k synthetically generated image–CXSMILES pairs from MarkushGrapher-Datasets (v1)
  • 91k samples from MolParser Dataset

Citation

If you use this dataset, please cite:

@inproceedings{strohmeyer2026markushgrapher2,
  title     = {MarkushGrapher-2: End-to-end Multimodal Recognition of Chemical Structures},
  author    = {Strohmeyer, Tim and Morin, Lucas and Meijer, Gerhard Ingmar and Weber, Valery and Nassar, Ahmed and Staar, Peter W. J.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

License

This Dataset is released under the Creative Commons Attribution 4.0 License.