Question–Answer (Q&A) pairs on rice for the development of AgriLLM

Ali, I.; Onaga, G.; Kouadio A.L.; Arouna, A.; Dossou-Yovo, Elliott Ronald; Ndindeng, S.A.; Gaudiose, M.; Twine, Edgar E.; Kalimuthu, S.; Zenna, N.; Amoah, N.K.A.; Semwal, V.K.; Elnamaky, R.; Tang, E.; Awio, T.

Question–Answer (Q&A) pairs on rice for the development of AgriLLM

Autores

Dossou-Yovo, Elliott Ronald

Editor

AfricaRice

Descripción

This document presents AfricaRice’s technical contribution to the AgriLLM initiative, launched by CGIAR and the UAE at COP29 (2024) to develop an open-source, agriculture-specific large language model tailored to smallholder realities in the Global South. The document emphasizes the importance of the post-training phase, where high-quality expert Q&A pairs are used both to fine-tune the model and to create evaluation benchmarks for iterative improvement. Within CGIAR, AgriLLM is positioned as a flagship product of the Action Lab under the CGIAR Digital Transformation Accelerator (DTA). AfricaRice reports a curated dataset of 531 expert-developed rice Q&A pairs, provided as a core resource for model training and evaluation. Methodologically, the dataset was produced through an expert-driven process led by AfricaRice scientists and senior staff across multiple disciplines (e.g., agronomy/soil science, breeding, pathology, grain quality and nutrition, and technology transfer). Each entry follows a structured template with standardized metadata such as unique ID, question, answer, language, target user persona, domain tags, and geography enabling systematic filtering, targeted fine-tuning, and robust evaluation workflows. In scope, the Q&A pairs reflect practical and decision-relevant questions commonly encountered in research and extension. Topics span cropping environments and stresses, seed systems and seed quality, planting density, fertilizer and water management (including AWD and mid-season drainage), varietal selection, and broader considerations such as grain quality, nutrition, and trade/compliance. Geographic references include Africa and other rice-growing regions, supporting context-specific and user-oriented model behavior.