Fine-tuning large language models for improved health communication in low-resource languages

Area of Science:

• Natural Language Processing
• Artificial Intelligence in Healthcare
• Low-Resource Language Technologies

Background:

• Healthcare information accessibility is limited in low-resource languages like Vietnamese.
• Large Language Models (LLMs) offer potential but require domain-specific adaptation.
• Developing countries face unique challenges in leveraging advanced AI for health.

Purpose of the Study:

• To develop a methodology for fine-tuning LLMs for Vietnamese healthcare information.
• To enhance medical communication and information accessibility in Vietnam.
• To adapt LLMs to specific linguistic and domain requirements of Vietnamese healthcare.

Main Methods:

• Selected three open-source LLMs as base models.
• Compiled a Vietnamese healthcare dataset of ~337,000 prompt-response pairs from various sources.
• Fine-tuned models using Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA) techniques.
• Evaluated performance using BertScore, Rouge-L, and LLM-as-a-Judge.

Main Results:

• Fine-tuned models showed significant performance improvements over base models across all metrics.
• Demonstrated the effectiveness of LoRA and QLoRA for domain-specific LLM adaptation.
• Highlighted the potential for improved healthcare communication in Vietnamese.

Conclusions:

• Fine-tuning LLMs is effective for low-resource languages like Vietnamese in healthcare.
• Significant computing power and costs present challenges for developing countries.
• Further initiatives are needed to promote global health equity through accessible AI.