PlantExpertVQA

A large-scale Visual Question Answering (VQA) dataset for plant disease diagnosis, built from 45 open-source image repositories. Questions span 9 categories across 3 cognitive complexity levels and are grounded in a structured Disease Knowledge Base of 203 expert-curated cards. Every QA pair underwent two-phase domain expert review by practising botanists.

Dataset at a Glance

Property	Value
Total QA pairs	765,186
Total images	150,841
Crop species	38
Disease conditions	89
KB cards	203
Question categories	9 (3 cognitive levels)
Splits	70 / 10 / 20 (train / val / test)
Image sources	45 open-source datasets
License	CC BY 4.0

Splits

Split	QA Pairs	Images
Train	535,881	105,586
Validation	76,384	15,080
Test	152,921	30,175

Splits are image-level stratified — no image appears in more than one split. Stratification is balanced across question category, crop species, disease condition, and source dataset.

Question Categories

Questions are organised into 3 cognitive levels following a Bloom-style hierarchy:

Level	Category	Description
L1 Foundational	Existence & Sanity Check	Confirms the image contains plant material
L1 Foundational	Plant Species Identification	Identifies the host crop
L1 Foundational	General Health Assessment	Healthy vs. diseased binary judgment
L2 Analysis	Visual Attribute Grounding	Detects specific visible symptoms
L2 Analysis	Detailed Verification	Verifies a named crop–disease combination
L3 Reasoning	Specific Disease Identification	Open-ended diagnosis without candidate
L3 Reasoning	Comprehensive Description	Full holistic description with severity
L3 Reasoning	Causal Reasoning	Identifies causal agent and risk factors
L3 Reasoning	Counterfactual Reasoning	Reasons about hypothetical healthy state

Data Schema

Each row in train.csv / val.csv / test.csv:

Column	Description
`qa_id`	Unique QA pair UUID
`image_id`	Filename of the image
`image_path`	Relative path to image (`images/<filename>`)
`crop`	Crop species (e.g., `tomato`, `apple`)
`disease`	Disease or condition label
`category`	Broad category: `disease`, `healthy`, or `senescence`
`severity`	`MILD` / `MODERATE` / `SEVERE` / `HEALTHY` / `UNKNOWN`
`question_text`	The question
`answer`	Ground-truth answer
`answer_type`	`open`, `closed`, or `free_form`
`question_category`	One of the 9 categories above
`cognitive_level`	`Level 1 (Foundational)` / `Level 2 (Analysis)` / `Level 3 (Reasoning)`
`dataset_source`	`PlantVQA` (KB-generated) or `PlantVillageVQA`

Disease Knowledge Base

disease_kb/ contains 203 JSON cards covering 38 crop species × 89 disease conditions. Each card encodes:

Pathogen taxonomy and transmission pathways
Organ-specific visual symptoms (leaves, stems, fruit, roots)
Severity rubric (MILD / MODERATE / SEVERE with quantitative thresholds)
Lookalike conditions with key differentiating features
Cultural, biological, and chemical management strategies

Cards are organised as disease_kb/<crop>/<disease>.json and aggregated in disease_kb/all_cards.jsonl.

Zero-Shot Benchmark Results

Evaluated on 250 images (1,178 QA pairs) from the test split. Primary metric: ROUGE-L.

Model	EM	Token-F1	BLEU-1	BLEU-2	ROUGE-L	BERTScore
Gemma-3-4B-IT	2.04	17.72	31.19	4.77	15.26	11.20
Qwen3-VL-2B-Instruct	3.29	17.88	37.18	6.63	14.54	8.53
Qwen2-VL-2B-Instruct	8.91	13.76	36.57	3.23	13.16	7.95
LLaVA-1.6 Mistral-7B	0.90	14.07	24.42	4.36	10.26	7.26
LLaVA-1.5-7B	7.39	10.21	26.87	0.58	9.87	3.15
InstructBLIP Vicuna-7B	6.54	9.70	28.53	2.18	9.40	−0.92
BLIP-2 FlanT5-XL	3.57	9.92	28.20	2.13	8.98	−2.74
CLIP ViT-L/14	3.31	7.44	15.30	2.27	6.95	3.48
BLIP VQA Large	5.43	6.65	21.41	0.04	6.64	−1.59

Fine-Tuning (Qwen3-VL-2B, LoRA, 500-image subset)

Setting	EM	Token-F1	BLEU-1	BLEU-2	ROUGE-L	BERTScore
Zero-shot best	2.04	17.72	31.19	4.77	15.26	11.20
Fine-tuned	18.85	64.65	66.18	41.74	61.70	63.27

Parameter-efficient LoRA fine-tuning on just 500 images (2,337 QA pairs) yields a 4× improvement in ROUGE-L over the best zero-shot model.

Usage

import pandas as pd

train = pd.read_csv("data/train.csv")
val   = pd.read_csv("data/val.csv")
test  = pd.read_csv("data/test.csv")

print(train.head())
print(train['question_category'].value_counts())

Using the datasets library:

from datasets import load_dataset

ds = load_dataset("SyedNazmusSakib/PlantExpertVQA")
print(ds['train'][0])

Repository Structure

PlantExpertVQA/
├── images_part1.zip  # 37,711 images (4.1 GB) — unzips to images/
├── images_part2.zip  # 37,711 images (4.1 GB) — unzips to images/
├── images_part3.zip  # 37,711 images (2.9 GB) — unzips to images/
├── images_part4.zip  # 37,708 images (0.6 GB) — unzips to images/
├── disease_kb/       # 203 JSON knowledge base cards
│   ├── all_cards.jsonl
│   └── <crop>/<disease>.json
└── data/
    ├── train.csv / train.jsonl   (535,881 QA pairs)
    ├── val.csv   / val.jsonl     ( 76,384 QA pairs)
    ├── test.csv  / test.jsonl    (152,921 QA pairs)
    └── split_statistics.json

Setup — Extracting Images

Download all four zip files and extract into the same directory:

unzip images_part1.zip
unzip images_part2.zip
unzip images_part3.zip
unzip images_part4.zip

All four zips unpack into the same images/ folder. The image_path column in the CSV files (images/<filename>) will then resolve correctly.

Citation

@article{sakib2025plantvillagevqa,
  title={PlantVillageVQA: 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}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence.
All constituent source datasets were released under permissive open licences (CC0, CC BY 4.0, or CC BY-NC 3.0). Full provenance for every image is preserved in the dataset metadata.

PlantExpertVQA

Dataset at a Glance

Property	Value
Total QA pairs	765,186
Total images	150,841
Crop species	38
Disease conditions	89
KB cards	203
Question categories	9 (3 cognitive levels)
Splits	70 / 10 / 20 (train / val / test)
Image sources	45 open-source datasets
License	CC BY 4.0

Splits

Split	QA Pairs	Images
Train	535,881	105,586
Validation	76,384	15,080
Test	152,921	30,175

Splits are image-level stratified — no image appears in more than one split. Stratification is balanced across question category, crop species, disease condition, and source dataset.

Question Categories

Questions are organised into 3 cognitive levels following a Bloom-style hierarchy:

Level	Category	Description
L1 Foundational	Existence & Sanity Check	Confirms the image contains plant material
L1 Foundational	Plant Species Identification	Identifies the host crop
L1 Foundational	General Health Assessment	Healthy vs. diseased binary judgment
L2 Analysis	Visual Attribute Grounding	Detects specific visible symptoms
L2 Analysis	Detailed Verification	Verifies a named crop–disease combination
L3 Reasoning	Specific Disease Identification	Open-ended diagnosis without candidate
L3 Reasoning	Comprehensive Description	Full holistic description with severity
L3 Reasoning	Causal Reasoning	Identifies causal agent and risk factors
L3 Reasoning	Counterfactual Reasoning	Reasons about hypothetical healthy state

Data Schema

Each row in train.csv / val.csv / test.csv:

Column	Description
`qa_id`	Unique QA pair UUID
`image_id`	Filename of the image
`image_path`	Relative path to image (`images/<filename>`)
`crop`	Crop species (e.g., `tomato`, `apple`)
`disease`	Disease or condition label
`category`	Broad category: `disease`, `healthy`, or `senescence`
`severity`	`MILD` / `MODERATE` / `SEVERE` / `HEALTHY` / `UNKNOWN`
`question_text`	The question
`answer`	Ground-truth answer
`answer_type`	`open`, `closed`, or `free_form`
`question_category`	One of the 9 categories above
`cognitive_level`	`Level 1 (Foundational)` / `Level 2 (Analysis)` / `Level 3 (Reasoning)`
`dataset_source`	`PlantVQA` (KB-generated) or `PlantVillageVQA`

Disease Knowledge Base

disease_kb/ contains 203 JSON cards covering 38 crop species × 89 disease conditions. Each card encodes:

Pathogen taxonomy and transmission pathways
Organ-specific visual symptoms (leaves, stems, fruit, roots)
Severity rubric (MILD / MODERATE / SEVERE with quantitative thresholds)
Lookalike conditions with key differentiating features
Cultural, biological, and chemical management strategies

Cards are organised as disease_kb/<crop>/<disease>.json and aggregated in disease_kb/all_cards.jsonl.

Zero-Shot Benchmark Results

Evaluated on 250 images (1,178 QA pairs) from the test split. Primary metric: ROUGE-L.

Model	EM	Token-F1	BLEU-1	BLEU-2	ROUGE-L	BERTScore
Gemma-3-4B-IT	2.04	17.72	31.19	4.77	15.26	11.20
Qwen3-VL-2B-Instruct	3.29	17.88	37.18	6.63	14.54	8.53
Qwen2-VL-2B-Instruct	8.91	13.76	36.57	3.23	13.16	7.95
LLaVA-1.6 Mistral-7B	0.90	14.07	24.42	4.36	10.26	7.26
LLaVA-1.5-7B	7.39	10.21	26.87	0.58	9.87	3.15
InstructBLIP Vicuna-7B	6.54	9.70	28.53	2.18	9.40	−0.92
BLIP-2 FlanT5-XL	3.57	9.92	28.20	2.13	8.98	−2.74
CLIP ViT-L/14	3.31	7.44	15.30	2.27	6.95	3.48
BLIP VQA Large	5.43	6.65	21.41	0.04	6.64	−1.59

Fine-Tuning (Qwen3-VL-2B, LoRA, 500-image subset)

Setting	EM	Token-F1	BLEU-1	BLEU-2	ROUGE-L	BERTScore
Zero-shot best	2.04	17.72	31.19	4.77	15.26	11.20
Fine-tuned	18.85	64.65	66.18	41.74	61.70	63.27

Parameter-efficient LoRA fine-tuning on just 500 images (2,337 QA pairs) yields a 4× improvement in ROUGE-L over the best zero-shot model.

Usage

import pandas as pd

train = pd.read_csv("data/train.csv")
val   = pd.read_csv("data/val.csv")
test  = pd.read_csv("data/test.csv")

print(train.head())
print(train['question_category'].value_counts())

Using the datasets library:

from datasets import load_dataset

ds = load_dataset("SyedNazmusSakib/PlantExpertVQA")
print(ds['train'][0])

Repository Structure

PlantExpertVQA/
├── images_part1.zip  # 37,711 images (4.1 GB) — unzips to images/
├── images_part2.zip  # 37,711 images (4.1 GB) — unzips to images/
├── images_part3.zip  # 37,711 images (2.9 GB) — unzips to images/
├── images_part4.zip  # 37,708 images (0.6 GB) — unzips to images/
├── disease_kb/       # 203 JSON knowledge base cards
│   ├── all_cards.jsonl
│   └── <crop>/<disease>.json
└── data/
    ├── train.csv / train.jsonl   (535,881 QA pairs)
    ├── val.csv   / val.jsonl     ( 76,384 QA pairs)
    ├── test.csv  / test.jsonl    (152,921 QA pairs)
    └── split_statistics.json

Setup — Extracting Images

Download all four zip files and extract into the same directory:

unzip images_part1.zip
unzip images_part2.zip
unzip images_part3.zip
unzip images_part4.zip

All four zips unpack into the same images/ folder. The image_path column in the CSV files (images/<filename>) will then resolve correctly.

Citation

@article{sakib2025plantvillagevqa,
  title={PlantVillageVQA: 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}
}

PlantExpertVQA

Dataset Card

PlantExpertVQA

Dataset at a Glance

Splits

Question Categories

Data Schema

Disease Knowledge Base

Zero-Shot Benchmark Results

Fine-Tuning (Qwen3-VL-2B, LoRA, 500-image subset)

Usage

Repository Structure

Setup — Extracting Images

Citation

License

Formats

License

Source

Paper for this model

Dataset Viewer

PlantExpertVQA

Dataset Card

PlantExpertVQA

Dataset at a Glance

Splits

Question Categories

Data Schema

Disease Knowledge Base

Zero-Shot Benchmark Results

Fine-Tuning (Qwen3-VL-2B, LoRA, 500-image subset)

Usage

Repository Structure

Setup — Extracting Images

Citation

License

Formats

License

Source

Paper for this model

Dataset Viewer