临床医学中大型语言模型的微调方法通过监督微调和直接偏好优化:比较评估
Thomas Savage1, Stephen P Ma2, Abdessalem Boukil3
1Division of Hospital Medicine, Perelman School of Medicine, Department of Medicine, University of Pennsylvania, 3400 Spruce St, Philadelphia, PA, 19147, United States, 1 2155191670.
Journal of medical Internet research
|September 23, 2025
概括
直接偏好优化 (DPO) 微调增强了大型语言模型 (LLM) 在临床推理和总结等复杂的医疗任务上的性能,而监督微调 (SFT) 则足以完成更简单的分类任务.
科学领域:
- 人工智能的人工智能
- 医疗信息学 医疗信息学
- 自然语言处理自然语言处理.
背景情况:
- 大型语言模型 (LLM) 在医学中显示出潜力,但往往需要微调才能达到最佳性能.
- 监督微调 (SFT) 和直接偏好优化 (DPO) 是关键的LLM微调技术.
- 在临床环境中,在SFT和DPO之间进行选择的指导是有限的.
研究的目的:
- 为了比较SFT和DPO对各种医疗自然语言任务的有效性.
- 为临床信息学家提供基于数据的建议,帮助他们选择微调方法.
- 为医疗保健运营中LLM的开发和部署提供信息.
主要方法:
- 使用Llama3 8B和Mistral 7B v2模型进行比较.
- 评估了SFT和DPO在四个核心医学NLP任务上的表现:文本分类,临床推理,总结和分类.
- 使用精度,利克特尺度和F1分数量化性能,以及计算资源分析.
主要成果:
- 与SFT相比,DPO显著改善了临床推理 (准确度从22%到40%) 和总结 (利克尔特尺度从3.93到4.08).
- SFT足以进行文本分类 (F1得分高达0.98),而DPO显示混合结果.
- 微调DPO需要比单独的SFT计算资源多2-3倍的计算资源.
结论:
- SFT适用于简单的任务,如文本分类.
- 在SFT后应用的DPO,提高了复杂任务的性能,包括分拣和临床推理.
- 这些发现指导了在医学应用中对LLM微调的战略部署.
相关概念视频
Improving Translational Accuracy
14.1K
Base complementarity between the three base pairs of mRNA codon and the tRNA anticodon is not a failsafe mechanism. Inaccuracies can range from a single mismatch to no correct base pairing at all. The free energy difference between the correct and nearly correct base pairs can be as small as 3 kcal/ mol. With complementarity being the only proofreading step, the estimated error frequency would be one wrong amino acid in every 100 amino acids incorporated. However, error frequencies observed in...
14.1K
Improving Translational Accuracy
3.6K
3.6K
Pharmacokinetic Models: Comparison and Selection Criterion
338
Physiological and compartmental models are valuable tools used in studying biological systems. These models rely on differential equations to maintain mass balance within the system, ensuring an accurate representation of the dynamic processes at play.
Physiological models take a detailed approach by considering specific molecular processes. They can predict drug distribution, metabolism, and elimination changes, providing a comprehensive understanding of how drugs interact with the body.
Physiological models take a detailed approach by considering specific molecular processes. They can predict drug distribution, metabolism, and elimination changes, providing a comprehensive understanding of how drugs interact with the body.
338
Mechanistic Models: Compartment Models in Algorithms for Numerical Problem Solving
292
Mechanistic models play a crucial role in algorithms for numerical problem-solving, particularly in nonlinear mixed effects modeling (NMEM). These models aim to minimize specific objective functions by evaluating various parameter estimates, leading to the development of systematic algorithms. In some cases, linearization techniques approximate the model using linear equations.
In individual population analyses, different algorithms are employed, such as Cauchy's method, which uses a...
In individual population analyses, different algorithms are employed, such as Cauchy's method, which uses a...
292
Dose-Response Relationship: Selectivity and Specificity
9.6K
Drugs exert their therapeutic effects by interacting with receptors, enzymes, or ion channels that are present throughout the human body. The strength and duration of the interaction between a drug and its target receptor are characterized by the selectivity and specificity of the drug. Selectivity refers to a drug's strong preference for its intended target over other targets. For instance, isoprenaline, a non-selective β-adrenergic agonist, interacts with both β1- and...
9.6K
Sensitivity, Specificity, and Predicted Value
1.2K
In healthcare diagnostics, laboratory tests play a crucial role in identifying and diagnosing a wide range of medical conditions. However, interpreting test results is not always straightforward. An abnormal test result does not always confirm the presence of a disease, just as a normal result does not guarantee its absence. To assess the reliability of these diagnostic tools, healthcare practitioners rely on two key statistical indicators: sensitivity and specificity.
Sensitivity is the...
Sensitivity is the...
1.2K

