CoolFace
Datasetpublic

toufiqmusah/GhanaAgricVQA-Dataset

GhanaAgricVQA — Crop Disease Visual Question Answering A Visual Question Answering (VQA) dataset for crop disease diagnosis in smallholder farming contexts, built from the RAIL Crop Disease Dataset.Each image has three QA pairs — identification, action, prevention — grounded in an expert knowledge base and written in conversational farmer-friendly English with Twi translations. Purpose: Fine-tune vision-language models (e.g., OWL-v2) to serve as AI crop disease assistants… See the full description on the dataset page: https://huggingface.co/datasets/toufiqmusah/GhanaAgricVQA-Dataset.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face
2likes106downloads
Dataset Card

GhanaAgricVQA — Crop Disease Visual Question Answering

<p align="center"> <img src="https://huggingface.co/datasets/GhanaAgricVQA-Dataset/resolve/main/assets/teaser.jpg" alt="GhanaAgricVQA samples" width="90%"/> </p>

A Visual Question Answering (VQA) dataset for crop disease diagnosis in smallholder farming contexts, built from the RAIL Crop Disease Dataset. Each image has three QA pairs — identification, action, prevention — grounded in an expert knowledge base and written in conversational farmer-friendly English with Twi translations.

Purpose: Fine-tune vision-language models (e.g., OWL-v2) to serve as AI crop disease assistants for smallholder farmers in sub-Saharan Africa.

Dataset Structure

SplitImagesQA PairsCrops
Train~670~2,010Maize, Pepper, Tomato
Test~117~351Maize, Pepper, Tomato
Total7872,361—

Each image contains 3 QA entries:

TypeQuestionAnswer
🧐 Identification"What is affecting my [crop]?"Identifies the disease with key symptoms
🛠️ Action"What can I do to treat it?"Practical step-by-step treatment advice
🛡️ Prevention"How do I prevent this in future?"Preventive measures for next seasons

Sample Entries

<p align="center"> <img src="https://huggingface.co/datasets/GhanaAgricVQA-Dataset/resolve/main/assets/samplepepper.jpg" alt="Pepper sample" width="30%"/> <img src="https://huggingface.co/datasets/GhanaAgricVQA-Dataset/resolve/main/assets/samplemaize.jpg" alt="Maize sample" width="30%"/> <img src="https://huggingface.co/datasets/GhanaAgricVQA-Dataset/resolve/main/assets/sample_tomato.jpg" alt="Tomato sample" width="30%"/> </p>

{
  "image": <PIL Image>,
  "question": "What disease is affecting my pepper plant?",
  "answer": "Your pepper plant has bacterial spot...",
  "question_type": "identification",
  "crop": "pepper",
  "language": "en",
  "question_tw": "Dɛn na ɛrewu me pepper dua no?",
  "answer_text_tw": "...",
  "rail_image_id": "id_3g2vbe",
  "generation_confidence": 0.87,
  "validated": false
}

Crops & Diseases Covered

🌽 Corn (Maize) — 8 classes

Gray Leaf Spot, Common Rust, Northern Leaf Blight, Maize Streak Virus, Cercospora Leaf Spot, Healthy

🌶️ Pepper — 10 classes

Bacterial Spot, Cercospora Leaf Spot, Early Blight, Fusarium Wilt, Late Blight, Phytophthora Leaf Blight, Leaf Curl Virus, Leaf Mosaic Virus, Septoria Leaf Spot, Healthy

🍅 Tomato — 8 classes

Bacterial Spot, Early Blight, Fusarium Wilt, Late Blight, Leaf Curl Virus, Mosaic Virus, Septoria Leaf Spot, Healthy

How the Data Was Created

  1. 1.Images: 4,902 RAIL crop disease images (originally from Ghana, West Africa)
  2. 2.Knowledge Base: Expert-synthesized disease information covering symptoms, treatments, and prevention for all 23 disease classes
  3. 3.Generation: Qwen3.5-9B via Ollama on NVIDIA L4 GPU
  4. 4.Each prompt injects the relevant KnowledgeBase section
  5. 5.2 parallel workers with checkpoint resume
  6. 6.Output structured JSON with Twi translations
  7. 7.Splits: 85/15 train-test random split (seed 42)

Usage

python
from datasets import load_dataset

dataset = load_dataset("GhanaAgricVQA-Dataset")

# Access training split
train = dataset["train"]
print(train[0]["question"]) 
# "What disease is affecting my maize plant?"

# Pair with images
image = train[0]["image"]
answer = train[0]["answer"]

Citation

bibtex
@misc{ghanaagrivqa2026,
  title={GhanaAgricVQA: Crop Disease Visual Question Answering and Bounding Box Detection Dataset},
  author={Toufiq Musah, Nyameye Akyaa Idun-Sam, Abotsi Benjamin Etornam},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/GhanaAgricVQA-Dataset}
}

License

This dataset is released under CC-BY 4.0. The underlying RAIL images retain their original license.