README.md

U-RxnDiagram-15k Dataset (Sci-Align)

🌌 The Sciverse Data Foundation

Sciverse is a comprehensive, multi-layered scientific data foundation designed to provide the ultimate data infrastructure for the AI for Science (AI4S) community. As scientific research becomes increasingly data-driven, Sciverse supplies the essential, high-quality data resources required to build robust scientific knowledge systems and accelerate research.

Sciverse consists of three core data pillars:

• Sci-Base (Scientific Knowledge Base Data): The massive-scale, purely objective scientific knowledge base. Comprising over 25 million deeply cleaned and parsed Open Access documents, it provides the comprehensive, purely factual scientific corpus that serves as the universal foundation for all downstream scientific applications.
• Sci-Align (Scientific Multi-Alignment Data): A highly curated, structured dataset mapping direct scientific relationships and precise factual alignments. It focuses on well-defined entity interactions—such as mapping specific chemical reaction pathways (e.g., via SMILES strings), condition-to-result pairings, and standardized structural descriptions. This layer provides the structured factual alignment needed for models to accurately connect and ground foundational scientific concepts.
• Sci-Evo (Scientific Evolution Data): A multi-layered, high-density reasoning dataset designed for complex problem-solving and deep scientific evaluation. Going beyond basic facts, this layer captures deep, causal descriptions—detailing not just the 'what', but the underlying reasoning for specific experimental designs, multi-step mathematical derivations, and the complex logic of how modifying specific conditions alters outcomes. It is constructed to rigorously measure a model's advanced scientific reasoning accuracy and logical depth.

---

U-RxnDiagram-15k Dataset Overview (Sci-Align)

U-RxnDiagram-15k Dataset is a large-scale dataset specifically designed for chemical reaction diagram parsing, containing chemical reaction images extracted from scientific literature PDFs along with detailed annotations. This dataset aims to support research in cheminformatics, document analysis, and computer vision fields.

Dataset Statistics

• Total Images: 15,400 images
  • Train set: 15,000 images
  • Test set: 400 images
• Total Reactions: 48,255 reactions
  • Train set: 45,426 reactions
  • Test set: 2,829 reactions
• Data Source: Scientific literature PDF files
• Image Format: PNG
• Total Annotations: Approximately 165,468 annotation instances

Dataset Structure

U-RxnDiagram-15k/
├── train_set/
│   ├── ground_truth.json          # Train set annotation file
│   └── images/                    # Train set image directory
└── test_set/
    ├── ground_truth.json          # Test set annotation file
    └── images/                    # Test set image directory

Annotation Category Definitions

The dataset defines 4 main categories, each containing multiple fine-grained attributes:

1. Structure - category_id: 1

• P-reactant: Reactant molecular structures
• P-product: Product molecular structures
• P-reaction condition: Reaction condition molecular structures

2. Text - category_id: 2

• T-reaction condition: Reaction condition text
• T-reactant: Reactant text
• T-product: Product text

3. Identifier - category_id: 3

• Chemical identifiers and numbers

4. Supplement - category_id: 4

• Other supplementary information

Annotation Statistics

Train Set

Attribute Type	Annotation Count	Percentage
T-reaction condition	56,377	35.84%
P-reactant	31,779	20.20%
P-product	30,808	28.79%
T-reactant	6,433	6.02%
T-product	3,804	2.42%
P-reaction condition	6,230	3.96%

Test Set

Attribute Type	Annotation Count	Percentage
T-reaction condition	3,011	36.92%
P-reactant	1,521	18.65%
P-product	2,348	28.79%
T-reactant	491	6.02%
T-product	388	4.76%
P-reaction condition	397	4.87%

Data Format

Image File Naming

Image filenames are hashed (SHA-256, first 8 hex chars). Example: a1b2c3d4.png.

Annotation File Format (ground_truth.json)

The annotation file follows COCO format and contains the following main fields:

{
  "licenses": [...],
  "info": {
    "description": "A dataset for chemical visual diagram analysis",
    "version": "v1",
    "year": "2025"
  },
  "categories": [
    {"id": 1, "name": "structure"},
    {"id": 2, "name": "text"},
    {"id": 3, "name": "identifier"},
    {"id": 4, "name": "supplement"}
  ],
  "images": [
    {
      "id": 2,
      "class": "figure",
      "confidence": 0.9148465991020203,
      "bbox": [x1, y1, x2, y2],
      "original_id": 0,
      "width": 1008.7104797363281,
      "height": 471.88232421875,
      "file_name": "ays765k9.png",
      "bboxes": [
        {
          "id": 0,
          "bbox": [x, y, width, height],
          "category_id": 1,
          "category": "P",
          "attribute": "P-reactants",
          "region_id": ["akzkPsql"]
        }
      ]
    }
  ]
}

Annotation Field Descriptions

• id: Unique annotation identifier
• bbox: Bounding box coordinates [x1, y1, x2, y2] for image-level bbox, [x, y, width, height] for bboxes list
• category_id: Category ID (1-4)
• category: Category abbreviation (P=Structure, T=Text)
• attribute: Specific attribute name
• region_id: List of region identifiers

Use Cases

This dataset is suitable for the following research tasks:

1. Chemical Structure Recognition: Identify and locate molecular structures in chemical reactions
2. Text Information Extraction: Extract text information from chemical diagrams
3. Reaction Condition Analysis: Identify and analyze reaction conditions
4. Document Understanding: Understand chemical information in scientific literature
5. Multimodal Learning: Combine visual and text information for chemical analysis

Data Quality

• All images are sourced from high-quality scientific literature
• Annotations are professionally verified for accuracy
• Contains samples of various chemical reaction types and complexities
• Supports fine-grained chemical information analysis

License

This dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

Terms of Use

You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.
• NonCommercial — You may not use the material for commercial purposes. Commercial use is prohibited without explicit permission from the licensor.

Citation

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

@article{song2025rxncaption,
  title={RxnCaption: Reformulating Reaction Diagram Parsing as Visual Prompt Guided Captioning},
  author={Song, Jiahe and Wang, Chuang and Jiang, Bowen and Wang, Yinfan and Zheng, Hao and Wei, Xingjian and Liu, Chengjin and Nie, Rui and Gao, Junyuan and Sun, Jiaxing and others},
  journal={arXiv preprint arXiv:2511.02384},
  year={2025}
}

Contact

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

---