CDDM (Crop Disease Domain Multimodal) Dataset

CDDM is a large-scale multimodal benchmark dataset built to advance vision-language models for crop disease diagnosis. It pairs 137,000 crop disease images with over 1 million instruction-following question-answer conversations covering disease identification, causes, symptoms, and prevention/treatment strategies.

This dataset is indexed on https://project-agml.github.io/ as part of the AgML python library. Standardized to the HF image_text_to_text format with a single conversational messages schema, converted to Parquet with image bytes embedded directly.

Dataset Construction

The image data was compiled from two sources:

Source	Images	Description
Web Data	62,000	Public agricultural datasets (Kaggle) plus web-crawled disease images
Private Data	75,000	Original images collected via field surveys across multiple farms and orchards

All images were annotated by agricultural experts with crop category, disease category, and appearance description. The dataset spans 16 crop categories and 60 crop disease categories; 48 categories contain 500+ images each, with the remaining 7 containing 200–500 images.

Two instruction-following data types were generated using GPT-4 prompting:

• Crop Disease Diagnosis QA — over 1 million multi-turn QA pairs per image, covering crop/disease identification, including deliberately-crafted negative-answer questions to counter models' tendency toward false-positive diagnoses. Avg. question length: 6.11 words; avg. answer length: 8.92 words.
• Crop Disease Knowledge QA — QA pairs generated from expert-curated disease knowledge text (symptoms, pathogen characteristics, transmission, prevention/control). Avg. question length: 9.69 words; avg. answer length: 130.41 words.

A held-out test set of 3,000 images (not included in training data) was used by the original authors for benchmark evaluation.

Usage

from datasets import load_dataset

ds = load_dataset("Project-AgML/CDDM")
first = ds["train"][0]

# Access an image — decoded to PIL automatically
img = first["images"][0]
img.show()

Schema

Every record shares the SAME columns so heterogeneous AgML datasets concatenate cleanly: id, images (embedded image bytes), messages, origin_dataset, and raw_metadata.

raw_metadata is a JSON-encoded string holding source fields not folded into messages (here: file_names pointing to the original image path, and annotated crop_category / disease_category where available); restore it with json.loads(row["raw_metadata"]). Image placeholders in messages align 1:1 with the images column.

Citation

@inproceedings{liu2024cddm,
      title={A Multimodal Benchmark Dataset and Model for Crop Disease Diagnosis},
      author={Liu, Xiang and Liu, Zhaoxiang and Hu, Huan and Chen, Zezhou and Wang, Kohou and Wang, Kai and Lian, Shiguo},
      booktitle={Computer Vision -- ECCV 2024},
      pages={157--170},
      year={2025},
      publisher={Springer Nature Switzerland},
      address={Cham},
      isbn={978-3-031-73016-0},
      doi={10.1007/978-3-031-73016-0_10}
}

Liu, Xiang; Liu, Zhaoxiang; Hu, Huan; Chen, Zezhou; Wang, Kohou; Wang, Kai; Lian, Shiguo (2025), "A Multimodal Benchmark Dataset and Model for Crop Disease Diagnosis", ECCV 2024, pp. 157-170

License

Released by the original authors (China Unicom AI Innovation Center) as an open-source initiative for agricultural multimodal research. Original source and download instructions: https://github.com/UnicomAI/UnicomBenchmark/tree/main/CDDMBench. This license information is for reference only and does not constitute legal advice — refer to the original repository for the authoritative license terms.