Towards building multilingual language model for medicine.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Medical Informatics

Background:

• Open-source, multilingual medical language models are crucial for global healthcare access.
• Existing models often lack multilingual capabilities and domain-specific training.
• Bridging the gap in linguistic diversity is essential for equitable medical information dissemination.

Purpose of the Study:

• To introduce a comprehensive multilingual medical corpus (MMedC) for domain adaptation.
• To propose a multilingual medical question-answering benchmark (MMedBench) for evaluating LLMs.
• To develop and assess advanced open-source multilingual medical large language models (LLMs).

Main Methods:

• Construction of MMedC: a 25.5B token corpus across 6 languages.
• Development of MMedBench: a multilingual, multi-choice QA benchmark with rationale.
• Evaluation of open-source LLMs, including models trained on MMedC, on MMedBench.

Main Results:

• MMed-Llama 3 (8B parameters) demonstrated superior performance on MMedBench and English benchmarks.
• The model achieved performance comparable to GPT-4 in medical QA tasks.
• Auto-regressive training on MMedC significantly enhanced LLM capabilities.

Conclusions:

• The presented corpus and benchmark facilitate the development of multilingual medical LLMs.
• MMed-Llama 3 represents a significant advancement in open-source medical AI.
• This work supports broader, equitable access to medical knowledge across diverse linguistic groups.