metadata
license: cc-by-nc-nd-4.0
RxnPatent Dataset
The RxnPatent dataset is specifically constructed for the task of chemical reaction flow diagram parsing. It comprises data extracted from World Intellectual Property Organization (WO) and United States (US) patents in chemistry, pharmaceuticals, and related fields.
Key Features
- Language Composition: The textual data primarily consists of Chinese and English, with a smaller number of samples in Japanese and Korean.
- Data Format: Follows the annotation format consistent with the RxnScribe project. Please refer to its documentation for details.
- Data Characteristics:
- Includes both single-step and multi-step reaction flow diagram images.
- The
diagram_typeof images is not annotated. - Contains negative samples (images with chemical structures that do not depict a reaction) to improve model discrimination.
- Annotation Quality: All data annotations were performed manually by professionals holding a Master’s degree in Chemistry, ensuring high accuracy and reliability.
Statistics
| Split | Total Images | Images with Reactions | Total Annotated Reactions |
|---|---|---|---|
| Train | 2,281 | 1,084 | 3,918 |
| Test | 162 | 84 | 272 |
File Structure
RxnPatent/
├── images/ # Directory containing all patent images
├── train.json # Annotation file for the training set
└── test.json # Annotation file for the test set