ScannedRxn

Dataset overview

ScannedRxn is a carefully curated, test-only benchmark for chemical reaction diagram parsing in historical scientific archives. It was introduced in Molecular Identifier Visual Prompt and Verifiable Reinforcement Learning for Chemical Reaction Diagram Parsing to evaluate model robustness and out-of-distribution generalization when digitizing legacy chemical knowledge.

Unlike benchmarks dominated by contemporary born-digital literature, ScannedRxn is collected from printed chemical publications spanning the 1950s to the 1990s. Its images reflect the visual distribution of early printed media: severe scanning noise, low contrast, typewriter-style fonts, irregular character spacing, and non-standard legacy layouts. These artifacts can obscure the boundaries between molecular structures, identifiers, and textual conditions, making ScannedRxn a challenging testbed for cross-era generalization.

The dataset contains 200 JPEG reaction diagrams, evenly divided among four topological complexities: single-line, multi-line, tree, and cyclic reactions, with 50 samples in each class. Every image includes instance-level bounding boxes, OCR text, component roles, reaction-region associations, and structured reaction relationships. There is no training split.

The release follows the annotation organization used by RxnCaption/U-RxnDiagram-15k.

Statistics

• Images: 200
• Reactions: 955
• Annotation instances: 1,930
• Image format: JPEG

Reaction diagram layout	Images
Single-line	50
Multi-line	50
Tree	50
Cyclic	50

Attribute	Category ID	Instances	Percentage
P-reactant	1	299	15.49%
P-product	1	615	31.87%
P-condition	1	59	3.06%
T-reactant	2	115	5.96%
T-product	2	163	8.45%
T-condition	2	679	35.18%

Directory structure

ScannedRxn/
├── README.md
├── README_ZH.md
├── ground_truth.json
└── images/
    └── <8-character SHA-256 prefix>.jpg

Annotation categories

The category definitions are compatible with U-RxnDiagram-15k:

1. structure (category_id: 1, abbreviation P)
   • P-reactant: reactant molecular structure
   • P-product: product molecular structure
   • P-condition: molecular structure used as a reaction condition
2. text (category_id: 2, abbreviation T)
   • T-reactant: reactant text
   • T-product: product text
   • T-condition: reaction-condition text

Data format

ground_truth.json is a COCO-style JSON object with the top-level fields licenses, info, categories, images, and roles. Each entry in images contains:

• id, original_id, and class: image identifiers and image class
• width and height: image dimensions in pixels
• bbox: image-level region in [x1, y1, x2, y2] format
• file_name: pseudonymized filename used in images/
• bboxes: annotated diagram elements
• reactions: reaction relationships between annotated elements
• reaction_type: one of single-line, multi-line, tree, or cyclic

Each item in bboxes contains an image-local id, an absolute box in [x, y, width, height] format, a normalized box in the same format, category fields, an attribute, region metadata, and OCR text. Structure instances normally have an empty text value.

Each reaction is represented as follows:

{
  "reactants": [0, 1],
  "conditions": [2],
  "products": [3]
}

The integers refer to bboxes[].id values within the same image.

Filename pseudonymization

Only image filenames and the matching images[].file_name values were renamed. For an original filename, including its extension, the new basename is calculated as:

digest = hashlib.sha256(original_filename.encode("utf-8")).hexdigest()[:8]
new_filename = digest + original_extension

License

This release is provided under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). Commercial use requires explicit permission from the rights holder. Users are responsible for complying with the terms applicable to source publications.

Citation

If you use this dataset in your research, please cite:

@article{song2026molecular,
  title={Molecular Identifier Visual Prompt and Verifiable Reinforcement Learning for Chemical Reaction Diagram Parsing},
  author={Song, Jiahe and Wang, Chuang and Wang, Yinfan and Zheng, Hao and Nie, Rui and Jiang, Bowen and Wei, Xingjian and Gao, Junyuan and Wang, Yubin and Wang, Bin and Wu, Lijun and Wu, Jiang and Yu, Qian and He, Conghui},
  journal={arXiv preprint arXiv:2603.15011},
  year={2026}
}

Contact

For questions or suggestions, contact songjiahe@pjlab.org.cn.