Dataset Extraction and Creation for Social Services in Africa

BAKOUBOLO ESSOWE JUSTIN, Catherine Nana Nyaah Essuman

Keywords: Social service, Language models, AI agents, Digital Services

TL;DR: This paper explores how fine-tuned open-source LLMs (like Llama, Gemma, and Deepseek) can be adapted for African contexts to power digital services in finance, healthcare, and communication

Abstract: Large Language Models (LLMs) are transforming digital services globally, yet their integration into localized and resource-constrained environments, such as those in Africa, remains underexplored. This paper presents an extensive approach from dataset creation to real-world model deployment of fine-tuned LLMs, including Llama, Gemma, and Deepseek for structured financial, healthcare, and communication services. We develop structured datasets tailored to African contexts, fine-tune several open-source models, and evaluate their ability to accurately extract key details from informal messages and convert them into structured JSON outputs. We integrate the best-performing model into our already existing WhatsApp-based AI assistant capable of performing tasks like sending reminders, scheduling payments, and providing healthcare reminders. A comparative analysis reveals the differences in model performances, highlighting the best approaches for efficient deployment in resource-limited African markets. Our findings suggest that LLM-based solutions are viable in bridging the digital services gap in low-resource settings, enabling inclusivity and accessibility.

Submission Number: 11