PlantMicro

PlantMicro is a benchmark for testing vision-language models on microscopic plant images. Most plant benchmarks use field or leaf photos, so the authors built one for the cell and subcellular level. It has 5,410 microscopy images and 9,718 question and answer pairs, covering 25 hosts, four biological domains (mycology, nematology, botany and cell biology) and three imaging methods (light, fluorescence and electron microscopy).

The images come from published datasets and open repositories, including fungal spores of grapevine, tomato and cucumber pathogens, plant-parasitic nematodes, potato tuber cells, pollen, and the Plant Organelles Database. The questions were made from each source's metadata, with wrong options taken from other values of the same field, nearby numbers for counting, and shifted boxes for localization. Two biologists checked the attributes and two domain experts reviewed all pairs.

The authors tested 13 models. Gemini 2.5 Flash had the best macro accuracy (63.08%) and Gemini 2.5 Pro the best micro accuracy (65.07%). The best open model, Qwen-VL 7B, reached 51.98%, and 68.63% with retrieval added. Pathogen classification was hard for every model: GPT-5 reached 34.93%, against a random baseline of 24.95%.

• Paper: Benchmarking Vision-Language Models for Microscopic Plant Image Understanding
• Original dataset: tqwei05/PlantMicro
• License: the authors do not state a license. The images come from other datasets, and the authors say they asked for permission where needed.

Tasks

Task	question_type	Rows	Format
Domain classification	domain	2,082	4 options
Modality identification	modality	2,047	4 options
Host identification	specie	1,369	4 options
Object localization	detection	1,083	Box as text
Pathogen classification	class	1,059	4 options
Stain or dye recognition	stain	771	4 options
Target counting	count_c	605	4 options
Organelle recognition	organell	540	4 options
Resin identification	resin	81	4 options
Organ prediction	organ	81	4 options

The question_type names are kept as written in the source, including specie and organell. Localization answers are pixel boxes written as x1,y1,x2,y2, as asked in the question.

Source group (dataset)	Images
pollen	1,688
nematode	816
grape	769
potato	626
podb_orgn	459
cucumber	423
tomato	348
barley	200
podb_e	81

Structure

train-0000-of-0001.parquet
images/
  image-train-000-of-001.zip
  path_to_shard.parquet

Each row has:

• images: list with one {"bytes": null, "path": <sha256>}. The path is the SHA256 of the image bytes and points into a zip shard through images/path_to_shard.parquet.
• id: plantmicro_train_<counter>, for example plantmicro_train_0.
• messages: one user turn (image, then question with options) and one assistant turn (answer).
• raw_metadata: JSON string with dataset, question_type and raw_image_paths.

Example:

{
  "images": [{"bytes": null, "path": "27955f406a2e6bab623961b619ccf5934a39c0f15d7ccafd56362eb5e0f73b4f"}],
  "id": "plantmicro_train_0",
  "messages": [
    {"role": "user", "content": [
      {"type": "image", "text": null},
      {"type": "text", "text": "What is the most likely microscopy modality used to acquire this image?\nA. light microscopy\nB. fluorescence microscopy\nC. electron microscopy\nD. none of above"}
    ]},
    {"role": "assistant", "content": [{"type": "text", "text": "A. light microscopy"}]}
  ],
  "raw_metadata": "{\"dataset\": \"cucumber\", \"question_type\": \"modality\", \"raw_image_paths\": [\"plantmicro/images/cucumber/powderymildew/cucumber_powderymildew_147.jpg\"]}"
}

Changes from the original data

• Built from VQAs.json, which lists each image with one to six questions. Each question becomes its own row, in source order, so all 9,718 questions are kept. The source has no split, so all rows are in one train split. The benchmark is meant for evaluation.
• Options come from the options field (choices for counting questions) and are added to the question as lines like A. text. All text is kept as written.
• Answers combine the option letter and its text, for example A. light microscopy. Counting questions use correct_choice_idx for the letter. Other questions match the answer to its option.
• In 2,587 questions (modality, resin and some stain questions) the answer is a list with one value. That value is used as the answer.
• raw_metadata keeps dataset and question_type. The image path as written is kept in raw_image_paths. The question, options, answer and correct_choice_idx are already in messages.
• All 5,410 image files are packed, and every one is used by a question. Images are stored once, deduplicated by SHA256, which leaves 5,369 unique images, because 37 groups of files in the source are byte-identical.

Usage

from agml import loadImageTextToTextDataset

dataset = loadImageTextToTextDataset("Project-AgML/PlantMicro")

sample = dataset["train"][0]
sample["images"]    # list of PIL images
sample["messages"]

Citation

@article{wei2026plantmicro,
  title={Benchmarking Vision-Language Models for Microscopic Plant Image Understanding},
  author={Wei, Tianqi and Yu, Xin and Chen, Zhi and Chapman, Scott and Huang, Zi},
  journal={arXiv preprint arXiv:2606.22497},
  year={2026}
}

This dataset is indexed and structured on https://project-agml.github.io/ as part of Project AgML @ UC Davis. 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.