🌿 Herb163CMR

A Fine-Grained Chinese Herbal Medicine Image–Text Retrieval Dataset

163 categories · 20,000 images · 60,000 image-text pairs · 3 descriptions per image

Overview • Statistics • Structure • Herb-LEGA • Citation

🌟 Overview

Herb163CMR is a fine-grained image–text retrieval dataset for Chinese herbal medicine.

The dataset contains 20,000 images from 163 herbal medicine categories. Each image is paired with three structured English morphological descriptions, resulting in 60,000 image–text pairs.

The three descriptions characterize complementary visual information at different granularities:

• S1: macroscopic form and overall appearance;
• S2: structural and cross-sectional characteristics;
• S3: fine-grained local morphological signs.

📊 Dataset Statistics

Split	Images	Descriptions
Train	15,995	47,985
Validation	2,028	6,084
Test	1,977	5,931
Total	20,000	60,000

📁 Dataset Structure

After extracting Herb163CMR.zip, the dataset is organized as follows:

Herb163CMR/
├── Herb163CMR_Descriptions.jsonl
├── splits/
│   └── seed_42/
│       ├── train.jsonl
│       ├── validation.jsonl
│       └── test.jsonl
├── category_001/
│   ├── image_001.jpg
│   └── ...
├── category_002/
│   └── ...
└── ...

Herb163CMR_Descriptions.jsonl contains the three structured morphological descriptions associated with each image.

The files under splits/seed_42/ define the training, validation, and test subsets.

🚀 Herb-LEGA

Herb163CMR is the benchmark dataset used by Herb-LEGA, a fine-grained image–text retrieval framework based on local feature enhancement and generative feature alignment.

Code: github.com/SimonWang555/Herb-LEGA

📝 Citation

If you use Herb163CMR in your research, please cite:

Herb163CMR: LLMs-Assisted Multi-Granularity Morphological Description for Fine-Grained Chinese Herbal Medicine Image–Text Retrieval.

The complete BibTeX entry will be added after publication.