🌾 East African Agricultural Multilingual Dataset + Gemma Fine-Tuned Model

This repository contains a high-quality multilingual agricultural instruction dataset designed for building AI systems that support smallholder farmers across East Africa.

It also demonstrates a fine-tuned Gemma 4 multilingual agricultural assistant trained on this dataset.

---

🌍 Supported Languages

• English
• Kiswahili
• Kinyarwanda
• Kikuyu (Gikuyu)
• Dholuo (Luo)
• Somali

---

🎯 Objective

This project addresses the lack of structured agricultural NLP datasets in under-resourced African languages.

It enables:

• Multilingual farmer advisory systems
• SMS-based agricultural assistants
• Crop and soil intelligence tools
• Rural AI accessibility solutions

---

🧠 Model: Gemma 4 Fine-Tuning

A Gemma 4 model (Google DeepMind) was fine-tuned using this dataset.

Fine-tuning Setup

Capabilities After Fine-Tuning

• Detects input language automatically
• Responds in the same language
• Provides better, contextual agricultural advice (soil, crops, pests, irrigation)
• Works in SMS-style short responses

---

🌱 Example Use Case

Input (Kiswahili):

Ni mbegu gani bora kwa maeneo yenye mvua kidogo?

Output: (Kiswahili)

Kwa eneo lenye mvua kidogo kama Machakos na maeneo ya ukame ya Kenya, mbegu bora zaidi ni zile zilizoidhinishwa kama zinazostahimili ukame. Kwa mahindi, aina kama Katumani Composite B (KCB), DH04 na DH02 hupendekezwa kwa sababu hukomaa haraka na huhitaji mvua kidogo. Kwa mtama, aina kama Gadam sorghum na Seredo zinafaa sana kwa ukame na pia zina soko zuri kwa chakula na biashara. Kwa uwele, Kakamega 1 ni aina inayostahimili hali kavu na hukua vizuri hata kwenye udongo duni. Kwa mikunde, kunde za aina ya KVU 27 na Machakos 66 ni nzuri kwa maeneo kavu, huku njegere za ICEAP 00557 zikijulikana kwa kustahimili ukame na kutoa mavuno mazuri. Kwa mazao ya mafuta, karanga (kama variety ya Chalimbana au JL 24) na ufuta (sesame) ni chaguo bora sana kwa sababu huhitaji maji kidogo sana. Mazao haya yakipandwa mapema mwanzoni mwa mvua na kwa kutumia mbinu za kuhifadhi unyevu wa udongo, huweza kutoa mavuno bora hata katika mazingira ya ukame.

---

🎥 Demo

A working prototype of the fine-tuned system is included below:

📹 Demo Video:

• **Demo Vid uploaded on YouTube **

This demo shows:

• Multilingual SMS interaction
• Real-time agricultural Q&A
• Automatic language switching
• Gemma 4 responses across African languages

---

📦 Dataset Format

Each sample contains the following fields: | Field | Description | |-------|-------------| | question | The agricultural question in the target language | | answers | A short, direct answer to the question | | enhanced_prompt | A richer, instruction-style rephrasing of the question | | enhanced_completion | A detailed, structured response with headings and explanations | json { "question": "Ubuhinzi bwo Gutera butavanga ubutaka ni ubuhe?", "answers": "Bugabanya ihungabana ry'ubutaka, bukabungabunga imiterere yabwo no kugabanya isatagurika ryabwo.", "enhanced_prompt": "Sobanura ubuhinzi bwo gutera butavanga ubutaka...", "enhanced_completion": "**Ubuhinzi bwo gutera butavanga ubutaka...." } The enhanced_prompt / enhanced_completion pairs are the primary fields used for Gemma 4 instruction fine-tuning.

🌾 Use Cases

• AI agricultural extension systems
• Farmer SMS advisory bots
• Climate-smart agriculture tools
• Rural multilingual AI assistants
• LLM fine-tuning for African languages

---

⚠️ Limitations

• Some synthetic translations may require human validation
• Dialect variations not fully covered
• Not a replacement for professional agronomists

---

🤝 Acknowledgement

This dataset was developed for the Uncharted Data Challenge, focusing on:

• Underserved agricultural domains
• Low-resource African languages
• Real-world AI impact

---

🙏 Credits & Adaptations

This project builds on and adapts the following resources:

• Dataset curation and multilingual adaptation supported by Adaptation Labs
• Gemma 4 by Google DeepMind — base model used for fine-tuning (model card)
• Agricultural knowledge and terminology adapted from publicly available extension resources by organizations including the FAO (Food and Agriculture Organization) and regional agricultural research institutes

---

📌 Citation

If you use this dataset or model, please cite:

@dataset{raynene2026eastafrica,
 author = {Ray Munene},
 title = {Agricultural QA Data for East Africa},
 year = {2026},
 publisher = {Hugging Face},
 url = {https://huggingface.co/datasets/RayNene/Agricultural-QA-Data-4-East-Africa}
}