enalis/LeafNet2.0

LeafNet 2.0

LeafNet 2.0 is a multiregional image–text dataset of plant leaf diseases for vision-language modeling. It compiles leaf images from openly licensed source datasets collected in many countries, harmonises their crop and disease labels, and pairs the images with literature-grounded, expert-validated captions.

Its companion benchmark is LeafBench 2.0, a multiple-choice VQA benchmark built from LeafNet 2.0 images.

• Paper: Nguyen, T. V., Nguyen Quoc, K., Harwath, D., Quach, L.-D., & Dao, P. D. (2026). A multiregional image-text dataset and benchmark for vision-language modeling of plant diseases.
• Collection: https://huggingface.co/collections/enalis/leafmd
• Code: https://github.com/EnalisUs/LeafBench

At a glance

Images	216,815 (224 × 224, .jpg / .png)
Crop–disease classes	190 (154 disease, 36 healthy)
Crops	38
Image–caption pairs	215,659 (image_caption_full.csv)
Expert-evaluated caption subset	6,015 (image_caption_eval.csv)
Images per class	18 – 8,336 (mean 1,141)

Dataset structure

leafnet2.0/
├── images/
│   ├── [crop]-[disease]/                      e.g. Coffee-Miner/
│   │   ├── [crop]-[disease]-id_00001.jpg
│   │   ├── [crop]-[disease]-id_00002.jpg
│   │   └── ...
│   └── ...
├── image_caption_pair/
│   ├── image_caption_full.csv
│   └── image_caption_eval.csv
├── metadata/
│   ├── literature_reference/
│   │   └── caption_reference.csv
│   └── country_info.csv
└── LICENSE.txt

images/

One folder per crop–disease class, named [crop]-[disease] (for example Coffee-Miner, Tomato-Late_blight). Each image is named [crop]-[disease]-id_[ID] with a five-digit identifier that is unique within its class.

image_caption_pair/image_caption_full.csv

Image paths and their captions.

Column	Description
image_path	Path of the image relative to the repository root
folder_name	Crop–disease class (same as the image folder)
caption	Image caption
legacy_file_name	File name of the image in the July 2026 release, for matching older results (empty for images added in October 2026)

All images of a class share one caption, built from the class's rows in caption_reference.csv. Images added in October 2026 (see Version history) use the caption of their class, or a caption built with the same template for new classes. The 1,156 images of Grape-Powdery_Mildew and Tomato-Gray_mold have no caption yet.

image_caption_pair/image_caption_eval.csv

The subset of captions used for evaluation, with the expert review flags: image_path, folder_name, caption, background_flag, crop_flag, disease_flag, primary_symptom_flag, disease_stage_flag, revision_flag, notes, legacy_file_name.

metadata/literature_reference/caption_reference.csv

The literature-based reference table that the captions were built on: one or two rows per class (early / late stage) with Folder Name, Crop, Disease, Pathogen, Scientific Name, Number (images in this release), Symptom Description, Main Symptom, Shape, Leaf Status, Background, Lesion information, leaf deformation, Hole, Fungal Growth, Presence of Exudate, Stage, Stage Description and Paper Link.

metadata/country_info.csv

Source datasets with their link, number of raw images (Raw) and country of collection.

Usage

from datasets import load_dataset

# Image–caption pairs and metadata tables
captions = load_dataset("enalis/LeafNet2.0", "image_caption_full", split="train")
reference = load_dataset("enalis/LeafNet2.0", "caption_reference", split="train")

# Images with class labels taken from the folder names
images = load_dataset("enalis/LeafNet2.0", data_dir="images", split="train")

To work with the files directly:

from huggingface_hub import snapshot_download
import pandas as pd
from PIL import Image

root = snapshot_download("enalis/LeafNet2.0", repo_type="dataset")
pairs = pd.read_csv(f"{root}/image_caption_pair/image_caption_full.csv")
row = pairs.iloc[0]
image = Image.open(f"{root}/{row.image_path}")
print(row.folder_name, row.caption)

License

The images are not relicensed under a single license: each image keeps the license of the source dataset it came from. The annotations created by the authors (captions, reference table, metadata) are released under CC BY 4.0. Source datasets, their licenses, the conditions that apply to users (attribution, non-commercial use of MangoLeafBD, share-alike for PlantVillage, GPL 2.0 for the Black Pepper images) and the changes made to the images are listed in LICENSE.txt.

Two datasets used in the study are not redistributed here because their terms do not permit it: Plant Pathology 2021 – FGVC8 and the Kaggle wheat leaf dataset. Obtain them from the original sources if needed.

Version history

• October 2026
  • Removed all images from Plant Pathology 2021 – FGVC8 (16,024 images: the Apple-Frog_eye_leaf_spot, Apple-Powdery_mildew and Apple-Rust classes and the FGVC8 part of Apple-Scab and Apple-Healthy) and from the Kaggle wheat leaf dataset (407 images, three Wheat-* classes).
  • Removed 35,388 byte-identical copies from Coffee-Healthy (18,983 images, 26 distinct) and Coffee-Miner (16,978 images, 547 distinct); one copy of each image is kept and the IDs of the two classes are renumbered. Rows of image_caption_eval.csv that referred to a removed copy now refer to the kept image, and 37 rows that then repeated another row's image were dropped.
  • Added 12,809 images: Citrus (11 classes) and Orange huanglongbing from the Citrus Disease Dataset, Grape (Downy Mildew, Esca, Healthy, Mites, new Powdery Mildew class) from the Grapevine Leaf Disease dataset, Pear (4 classes) from DiaMOS, Tomato Gray mold, and Tea images from the Tea Leaf Disease dataset. New images were resized to 224 × 224 like the rest of the dataset, and 11,653 of them have captions.
  • Reorganised the repository into image folders and CSV files, and added LICENSE.txt.
• July 2026 – first release (Parquet), 255,825 images in 179 classes.

Citation

@misc{nguyen2026leafnet2,
  title  = {A multiregional image-text dataset and benchmark for vision-language modeling of plant diseases},
  author = {Nguyen, Trang V. and Nguyen Quoc, Khang and Harwath, David and Quach, Luyl-Da and Dao, Phuong D.},
  year   = {2026},
  doi    = {10.57967/hf/10895}
}

Contact

Trang V. Nguyen (tvn524@eid.utexas.edu) · Phuong D. Dao (phuong.dao@austin.utexas.edu) · or open an issue at https://github.com/EnalisUs/LeafBench