Efficient fine-tuning of small-parameter large language models for biomedical bilingual multi-task applications

Abstract

The escalating computational costs of large language models (LLMs) have catalyzed the pursuit of more efficient alternatives, particularly in specialized domains like biomedicine. In this study, we propose BioQwen, a series of small-parameter biomedical bilingual (Chinese–English) multi-task models designed to mitigate the resource demands of LLMs while achieving high performance.

BioQwen is trained on carefully curated open-source biomedical datasets, employing a stringent preprocessing pipeline with thorough quality filtering and standardized formatting. Through an efficient two-stage fine-tuning strategy, BioQwen models with 0.5B, 1.5B, and 1.8B parameters attain competitive performance across a variety of comprehension and generative tasks. For comprehension tasks, BioQwen-1.8B achieves a Macro F1 score of 0.730 and a balanced accuracy of 0.802 on the BC5CDR dataset, surpassing the 7B-parameter Taiyi model’s scores of 0.685 and 0.757. In generative tasks, BioQwen delivers superior zero-shot results on the iCliniq dataset, outperforming all baselines across multiple metrics. Comparisons with established small-parameter LLMs (e.g., Llama3.2 1B) further substantiate the effectiveness of domain-specific fine-tuning.

Significantly, BioQwen’s reduced iteration time highlights its computational efficiency, and its quantized version demonstrates successful deployment on mobile devices, confirming its viability in resource-constrained settings. This study demonstrates the potential of strategically fine-tuned small-parameter LLMs to deliver resource-efficient, high-performing solutions for biomedical bilingual applications, expanding accessibility and usability in the field.