GrowMate: Hydroponic Farming Chatbot for Rwanda

A domain-specific chatbot built with FLAN-T5-base, fine-tuned on 625 hydroponic farming Q&A pairs to provide expert guidance for smallholder farmers in Rwanda.

Links & Resources

• Live Demo: Try GrowMate
• GitHub: Source Code
• Hugging Face: Fine-tuned Model
• Demo Video: YouTube

About

GrowMate addresses the information gap in hydroponic farming for Rwandan farmers. Hydroponics is a soil-free, water-efficient method that uses 90% less water than traditional farming. This chatbot provides instant, accessible guidance on system setup, nutrient management, pH control, and troubleshooting.

Features

• FLAN-T5-base (247M parameters) fine-tuned on 625 Q&A pairs
• 283% ROUGE-2 improvement through 3 systematic experiments
• Live web interface with Streamlit
• Rwanda-specific context and farming conditions
• Complete metrics: ROUGE, BLEU, F1, Perplexity

Repository Structure

Farmsmart_growmate_chatbot/
├── .streamlit/config.toml          # Streamlit configuration
├── Assets/
│   ├── loader.png                  # Bot avatar
│   └── logo.png                    # App logo
├── data/
│   └── hydroponic_FAQS.csv         # Training dataset
├── notebooks/
│   └── flan_t5_hydroponic_chatbot.ipynb  # Model training
├── trained_model/                  # Fine-tuned model files
├── app.py                          # Streamlit app
├── requirements.txt                # Dependencies
└── README.md

Quick Start

1. Environment Setup

cd Farmsmart_growmate_chatbot
pip install -r requirements.txt

2. Train the Model

Open and run notebooks/flan_t5_hydroponic_chatbot.ipynb to train the model. The trained model saves to trained_model/ automatically.

3. Launch the Chatbot

streamlit run app.py  # Opens at http://localhost:8501

Model Training

The notebook covers:

• Data loading and cleaning
• FLAN-T5-base model setup
• Data preprocessing for instruction tuning
• Model training and evaluation
• Model saving for deployment

Experiment Results

Experiment Tracking

Performance Improvements

• 283% ROUGE-2 improvement (0.0125 → 0.0479)
• 42.1% training loss reduction (4.1165 → 2.3853)
• 48.6% ROUGE-1 improvement (0.1387 → 0.2061)
• No overfitting - consistent improvement across experiments

Additional Metrics (Experiment 3)

• BLEU: 0.0116
• F1 Score: 0.1357 (Precision: 0.1270, Recall: 0.1510)
• Perplexity: 1.3675 (excellent for domain-specific models)

Evaluation Metrics

ROUGE Scores:

• ROUGE-1: Unigram overlap
• ROUGE-2: Bigram overlap (semantic coherence)
• ROUGE-L: Longest common subsequence

BLEU Score: N-gram precision between generated and reference text

F1 Score: Harmonic mean of precision and recall

Perplexity: Model confidence (lower is better, <10 is very good)

Sample Model Predictions

Example 1:

• Q: Can I grow dill hydroponically?
• Model: Dill is a perennial herb that grows well in hydroponic systems; it can be grown as a seedling or in small containers.

Example 2:

• Q: Why are my roots brown?
• Model: Brown roots indicate nutrient deficiency or poor water quality. Check pH and airflow to ensure adequate nutrients are in the system.

Example 3:

• Q: What pH level should I maintain for hydroponic tomatoes?
• Model: Maintain pH 5.8–6.2; maintain EC 1.3–1.8; avoid overwatering or nutrient deficiency in tomatoes.

Usage Example

streamlit run app.py  # Opens at http://localhost:8501

Example Questions:

• "What pH level is best for lettuce?"
• "How often should I change nutrient solution?"
• "Best hydroponic crops for Rwanda?"

System Requirements

• Minimum: 4GB RAM, CPU-only inference
• Recommended: 8GB RAM, GPU for training
• Response time: 1-3 seconds per query

Deployment

Deployed on Streamlit Cloud: https://farmsmartgrowmatechatbot.streamlit.app/

The chatbot is a single-file Streamlit app with automatic model loading.

Future Work

• Immediate: Kinyarwanda language support, mobile app
• Advanced: IoT sensor integration, multimodal capabilities (image-based disease diagnosis)
• Long-term: Regional expansion, field validation with farmers

Contributing

Open source contributions welcome! For technical help or collaboration opportunities, see the GitHub repository.

---