| Summary: | This study explores how integration of domain-adapted Large Language Models (LLMs) and existing data-driven frameworks enhance wheat agronomic advisory services in east Africa. Although LLMs hold transformative potential across various domains, their use in agro-advisory services in resource-constrained regions like East Africa is limited by a lack of domain-specific data and inadequate infrastructure. The general objective of this study is develop domain-adapted large language models (LLMs) that integrate heterogeneous agronomic inputs with existing data-driven frameworks, and evaluate their performance, and cross-regional adaptability for delivering accurate and context-aware agro-advisories. By extracting and structuring heterogeneous inputs from diverse sources using LLMs an agronomic knowledge base will be developed. Then open-source LLMs via efficient techniques (e.g., QLoRA), will be fine-tuned and integrated with existing geospatially enabled advisory systems using Retrieval-Augmented Generation (RAG). Performance of LLMs will be assessed through quantitative metrics and expert validation. Cross-regional adaptability will be also tested by using in-context learning and evaluated using expert validation. Expected outcomes include efficient pipeline that extracts and curates agronomic information from diverse sources, alongside a domain-adapted LLM integrated with existing data‐driven frameworks. Finally findings of the study will be disseminated via publications and presentations
|
|---|