Multi-Dimensional Analysis of LLM Models with Synthetic Data on Early Marriage Prevention

Early marriage is a serious social issue that impacts the well-being of society, with approximately 12 million girls marrying before the age of 18 each year according to UNICEF data. This study aims to analyze the performance of three Large Language Models (LLMs) - Mistral-7B-Instruct, Llama-3.1-SB-Instruct, and Qwen2.5-7B-Instruct - using synthetic data in the context of early marriage prevention. The models were developed through fine-tuning and Retrieval- Augmented Generation (RAG) to improve text relevance and accuracy, and evaluated using UniEval. The evaluation results showed that Qwen2.5-7B obtained the highest score of 3.893, Llama-3.1-8B showed strength in coherence (up to 0.998) and groundedness (0.999) and Mistral-7B excelled in naturalness improvement after fine-tuning (0.900). This study highlights that the balance between readability, factual accuracy, and conversational appeal is key in improving the output quality of AI assistants for social education on early marriage prevention.