| --- |
| pretty_name: LeafNet |
| dataset_info: |
| - config_name: train |
| splits: |
| - name: train |
| num_examples: 121337 |
| configs: |
| - config_name: train |
| data_files: |
| - split: train |
| path: train-*.parquet |
| license: cc-by-4.0 |
| task_categories: |
| - image-text-to-text |
| - visual-question-answering |
| language: |
| - en |
| tags: |
| - vqa |
| - agriculture |
| - vision-language |
| - computer-vision |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| # LeafNet |
|
|
| LeafNet is a large-scale multimodal dataset for plant disease diagnosis, |
| introduced in ["LeafNet: A Large-Scale Dataset and Comprehensive Benchmark |
| for Foundational Vision-Language Understanding of Plant Diseases"](https://arxiv.org/abs/2602.13662) |
| (arXiv:2602.13662). The full dataset comprises 186,000 leaf images across |
| 22 crop species and 97 classes (43 fungal diseases, 8 bacterial diseases, |
| 2 mould/oomycete diseases, 6 viral diseases, 3 mite-induced diseases, plus |
| healthy leaves), each paired with an expert-curated symptom description. |
| The accompanying LeafBench VQA benchmark evaluates models on six tasks |
| (crop identification, healthy/diseased classification, disease |
| identification, symptom recognition, pathogen classification, and |
| scientific nomenclature); closed-source models such as GPT-4o reached up |
| to 72% accuracy, while the domain fine-tuned SCOLD model reached 99.15% on |
| disease identification. |
|
|
| ## Notes |
| The public dataset is just ~70% subset of the full 186,000-image dataset (the remainder is held out |
| by the original authors). |
|
|
| ## Layout |
|
|
| Single config (`train`), matching every other AgML-standardized dataset's |
| storage format with metadata parquet at the repo root, image shards under |
| `images/`: |
|
|
| ``` |
| LeafNet_P/ |
| train-0000-of-0001.parquet # 121,337 rows: images, id, messages, raw_metadata |
| images/ |
| image-train-000-of-003.zip |
| image-train-001-of-003.zip |
| image-train-002-of-003.zip |
| path_to_shard.parquet # 121,337 rows: path -> shard_file |
| ``` |
|
|
| `images` holds `{"bytes": None, "path": ...}` per row's single image. |
| `raw_metadata` keeps the original source `file_name` path for provenance, the verbatim `caption`. |
| `messages` is a single-turn conversion, since the source data is an |
| image-captioning dataset (one caption per image, no original question |
| field): a fixed instruction prompt (mentioned below) asking the model to describe the leaf's |
| condition, with the original caption as the assistant's answer was used to convert it to conversation. |
| ``` |
| USER_PROMPT = "Describe the condition of this plant leaf, including any visible disease and its symptoms." |
| ``` |
|
|
| ## Usage |
|
|
| ```python |
| from agml import loadImageTextToTextDataset |
| |
| ds, store = loadImageTextToTextDataset("Project-AgML/LeafNet", token=HF_TOKEN) |
| print(ds) # DatasetDict({'train': ...}) |
| ds["train"][0] # images decoded lazily on access |
| ``` |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite the original LeafNet paper: |
|
|
| ```bibtex |
| @misc{nguyenquoc2026leafnet, |
| title = {LeafNet: A Large-Scale Dataset and Comprehensive Benchmark for Foundational Vision-Language Understanding of Plant Diseases}, |
| author = {Nguyen Quoc, Khang and Dao, Phuong D. and Quach, Luyl-Da}, |
| year = {2026}, |
| eprint = {2602.13662}, |
| archivePrefix = {arXiv}, |
| primaryClass = {cs.CV}, |
| url = {https://arxiv.org/abs/2602.13662} |
| } |
| ``` |
|
|
| The dataset also underlies the SCOLD vision-language model: |
|
|
| ```bibtex |
| @article{NGUYENQUOC2025130084, |
| title = {A Vision-Language Foundation Model for Leaf Disease Identification}, |
| journal = {Expert Systems with Applications}, |
| pages = {130084}, |
| year = {2025}, |
| issn = {0957-4174}, |
| doi = {https://doi.org/10.1016/j.eswa.2025.130084}, |
| author = {Khang {Nguyen Quoc} and Lan Le {Thi Thu} and Luyl-Da Quach}, |
| } |
| ``` |
|
|
| --- |
|
|
| This dataset is indexed and structured on https://project-agml.github.io/ as part of the AgML python library. This dataset was reformatted from its original format to match HuggingFace's Imagefolder standards but requires an external module (agml) that processes and returns a HF Dataset object faster than HF module functions. |
