•  
  •  
 

Abstract

Sexual and reproductive health (SRH) education remains a sensitive and underserved domain in Indonesia, where sociocultural factors often constrain access to medically reliable information. While large language models (LLMs) enable conversational health assistance, general-purpose models lack explicit domain control, raising safety and reliability concerns in sensitive medical contexts. This study presents a domain-specific SRH conversational system developed through supervised instruction tuning of LLaMA-2-7B-Chat using 93,299 instruction-formatted samples derived from physician-authored question–answer interactions. The training objective follows causal language modeling over structured instruction sequences. At the same time, domain boundaries are enforced through explicit refusal modeling embedded directly in the instruction-tuning data, supporting the model in declining out-of-domain queries without relying solely on post hoc filtering. Parameter-efficient adaptation is implemented using QLoRA to support scalable training under limited computational resources. Evaluated on an independent held-out test set under identical decoding settings, the fine-tuned model improves over the base LLaMA-2-7B-Chat model in ROUGE-L (0.109 vs. 0.019), METEOR (0.096 vs. 0.048), and BERTScore (0.813 vs. 0.741), while BLEU remains low for both models (0.68 vs. 0.70); these gains reflect stronger content and semantic alignment rather than verbatim overlap. Human-centered evaluation (n=12) further indicates high perceived usefulness (4.78/5) and consistent domain-boundary enforcement (5.00/5 refusal correctness). These results suggest that large-scale, medically grounded instruction tuning combined with explicit refusal-based domain modeling provides an effective approach for developing reliable, domain-constrained conversational AI systems for SRH education.

First Page

45

Last Page

54

Download

Share

COinS