Taiyi: a bilingual fine-tuned large language model for diverse biomedical tasks.

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Artificial Intelligence in Healthcare
• Computational Linguistics

Background:

• Existing biomedical large language models (LLMs) primarily focus on monolingual question answering and conversation.
• There is a need to evaluate LLM performance across diverse biomedical NLP tasks and languages.

Purpose of the Study:

• To develop and evaluate Taiyi, a bilingual fine-tuned LLM for a wide range of biomedical NLP tasks.
• To assess the effectiveness of supervised fine-tuning strategies on diverse biomedical datasets.

Main Methods:

• Curated 140 biomedical text mining datasets (102 English, 38 Chinese) across 10+ task types.
• Converted corpora into instruction data for supervised fine-tuning of a general LLM.
• Employed a 2-stage fine-tuning strategy to optimize performance.

Main Results:

• Taiyi achieved superior performance on 13 test sets, including named entity recognition, relation extraction, text classification, and question answering.
• Demonstrated considerable potential for bilingual biomedical multitasking in a case study.
• Outperformed general LLMs on various biomedical NLP benchmarks.

Conclusions:

• High-quality biomedical corpora and effective fine-tuning strategies significantly enhance LLM performance in the biomedical domain.
• Taiyi exhibits bilingual multitasking capabilities via supervised fine-tuning.
• Generative LLMs still face challenges in non-generation tasks like information extraction, where discriminative models remain superior.