brimbim/agriculture-dataset

Agriculture Crop Disease Advisor Dataset v2

High-quality instruction dataset used to train the v2 model that achieved 92% win rate in the Adaption Labs AutoScientist Challenge (Agriculture Category).

Dataset Details

• Version: v2 (Round 2)
• Size: ~20,000+ examples (PlantVillage + 6,000+ synthetic)
• Format: Instruction-completion pairs (JSONL)
• Quality Score (Adaption): 9.8/10 (Grade A)
• Created for: Adaption Labs AutoScientist Challenge – Agriculture (Part 2)

Composition

• Source Data: PlantVillage dataset (leaf images across 38 disease classes)
• Synthetic Data: 6,000+ diverse examples covering real-world smallholder scenarios, multiple regions (East Africa, South Asia, Latin America, etc.), climate conditions, and practical farming constraints
• Focus: Smallholder farmers in developing regions with emphasis on accessibility, sustainability, and low-cost solutions

Example Format

{
  "instruction": "You are an expert agronomist helping smallholder farmers. A farmer in East Africa reports yellow spots with white centers on tomato leaves after heavy rain.",
  "output": "**Diagnosis**: Likely late blight on tomato.\n\n**Recommended Actions**:\n• Remove and destroy affected leaves\n• Apply organic fungicide (e.g. neem-based)\n• Improve airflow and avoid overhead watering\n\n**Confidence**: 85%\n**Risk**: High - can spread quickly\n**Prevention**: Use resistant varieties and proper spacing."
}

Intended Use

• Fine-tuning agriculture-focused LLMs
• Building practical farmer advisory tools
• Research on domain adaptation for real-world impact (food security, climate resilience)

Model Trained With This Dataset

Agriculture Crop Disease Advisor v2 (92% win rate) → https://huggingface.co/brimbim/agriculture-crop-disease-model-v2

License

Apache 2.0 – Free to use for research and commercial purposes with attribution.

Citation

bibtex

@misc{agriculture-crop-disease-dataset-v2, author = {brimbim}, title = {Agriculture Crop Disease Advisor Dataset v2}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/datasets/brimbim/agriculture-dataset}}, note = {Created for Adaption Labs AutoScientist Challenge} }