PlantExpertVQA

PlantExpertVQA is a visual question answering dataset for plant disease diagnosis. It has 765,186 questions and answers over 150,841 leaf images from 45 open datasets, covering 38 crops and 89 disease conditions. The questions fall into 9 categories at 3 levels: identification, symptom checking and reasoning.

The questions come from templates filled from a knowledge base of 203 expert-written disease cards, then reworded by hand. No free-form LLM generation was used. Botanists reviewed the data in two rounds.

The authors tested 9 models zero-shot. The best, Gemma-3-4B-IT, reached a ROUGE-L of 15.26. Fine-tuning Qwen3-VL-2B with LoRA on just 500 images raised it to 61.70.

All 193,609 rows of PlantVillageVQA are part of this dataset (dataset_source is PlantVillageVQA for them). The other 571,577 rows are new (PlantVQA).

• Paper: PlantExpertVQA: A Visual Question Answering Dataset for Benchmarking Vision-Language Models in Plant Science, accepted at Scientific Data
• Original dataset: SyedNazmusSakib/PlantExpertVQA

Question categories

Level	Category	Train	Validation	Test
1	Existence & Sanity Check	1,950	258	637
1	Plant Species Identification	75,920	10,832	21,695
1	General Health Assessment	46,656	6,631	13,285
2	Visual Attribute Grounding	25,403	3,662	7,167
2	Detailed Verification	59,888	8,596	17,162
3	Specific Disease Identification	139,248	19,942	39,811
3	Comprehensive Description	62,550	8,885	17,829
3	Causal Reasoning	77,643	11,075	22,166
3	Counterfactual Reasoning	43,399	6,121	12,261
	Empty rows (no question or answer)	3,224	382	908
	Total	535,881	76,384	152,921

The splits use 105,586, 15,080 and 30,175 images. No image is in more than one split.

Structure

train-0000-of-0001.parquet
validation-0000-of-0001.parquet
test-0000-of-0001.parquet
images/
  image-train-000-of-NNN.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: plantexpertvqa_<split>_<counter>, for example plantexpertvqa_validation_0.
• messages: one user turn (image, then question) and one assistant turn (answer).
• raw_metadata: JSON string with qa_id, image_id, crop, disease, category, severity, answer_type, question_category, cognitive_level, dataset_source and raw_image_paths.

Example:

{
  "images": [{"bytes": null, "path": "ff09cdfc69def263fd7282051dc059943e66f631f47d1765ee5ddd975d981cf5"}],
  "id": "plantexpertvqa_validation_0",
  "messages": [
    {"role": "user", "content": [
      {"type": "image", "text": null},
      {"type": "text", "text": "What disease is shown here?"}
    ]},
    {"role": "assistant", "content": [{"type": "text", "text": "This is gourd anthracnose, displaying severe symptoms including extensive water-soaked spots and tissue decay."}]}
  ],
  "raw_metadata": "{\"qa_id\": \"105fe473-2b9c-4730-b34e-4d584ea7deac\", \"image_id\": \"fdf20db891df9a3e.jpg\", \"crop\": \"bottle gourd\", \"disease\": \"gourd anthracnose\", \"category\": \"disease\", \"severity\": \"SEVERE\", \"answer_type\": \"\", \"question_category\": \"Specific Disease Identification\", \"cognitive_level\": \"Level 3 (Reasoning)\", \"dataset_source\": \"PlantVQA\", \"raw_image_paths\": [\"images/fdf20db891df9a3e.jpg\"]}"
}

Changes from the original data

• Built from data/train.parquet, data/val.parquet and data/test.parquet. The CSV and JSONL files hold the same rows. val.parquet becomes the validation split, as named in the source card.
• Questions and answers are used exactly as in the source. There are no answer options.
• 4,514 rows have an empty question, answer, category and level in the source. They all belong to 680 images that have no other rows. They are kept as they are, so their user and assistant text is empty.
• raw_metadata keeps the source columns except question_text, answer and image_path, which are already in messages. Empty values stay empty, for example answer_type is empty in 520,843 rows.
• All 150,841 source images are packed. Every one is used by at least one row. Images are stored once, deduplicated by SHA256.
• The disease_kb/ knowledge base cards are also a part of this repo.

Usage

from agml import loadImageTextToTextDataset

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

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

Citation

@article{sakib2025plantexpertvqa,
  title={PlantExpertVQA: A Visual Question Answering Dataset for Benchmarking Vision-Language Models in Plant Science},
  author={Sakib, Syed Nazmus and Haque, Nafiul and Hossain, Mohammad Zabed and Arman, Shifat E},
  journal={arXiv preprint arXiv:2508.17117},
  year={2025}
}

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.