Papers

S-BDS 双层数据选择方法在医疗 LLM 微调中兼顾安全与能力

Safety-Preserving Medical LLM Fine-Tuning via Bilevel Data Selection

Why we picked it

做医疗方向微调的可以看看,论文给了具体的双层数据筛选做法,在 Llama 和 Qwen 上安全率比最强基线高出 4.64 和 3.00 个百分点,还提供了能力与安全之间的权重调节手段,不用全量人工标注也能控风险。

arXiv 论文 2611.20643 提出 Scalable Bilevel Data Selection(S-BDS),用于面向患者的医疗大模型监督微调(SFT)。方法通过控制选择权重的集中度,在大规模候选池中稳定排序样本,筛出既能提升医疗能力又不破坏安全对齐的训练数据,并将临床医生审核过的安全数据与筛选数据混合训练。

Body

在 HealthBench 和 PatientSafetyBench 两个基准上,实验覆盖 Llama 与 Qwen 两类模型,S-BDS 在所有对比方法中取得最高的安全率,分别超过最强基线 4.64 和 3.00 个百分点,同时医疗能力保持竞争力。论文还显示,调节 SFT 中安全导向权重可以控制医疗能力与患者安全的取舍曲线。

对做医疗垂直微调的团队来说,这项工作提供了一条可操作的数据筛选路径:不必全量人工审核指令数据,也能在引入领域知识的同时守住安全底线。

Original · arXiv cs.AI

Safety-Preserving Medical LLM Fine-Tuning via Bilevel Data Selection

As general large language models (LLMs) are increasingly adapted for patient-facing medical applications, improving medical capability while controlling patient-facing risks remains a critical challenge. Large-scale medical instruction data are essential for acquiring domain knowledge through supervised fine-tuning (SFT), but unsafe responses can undermine existing safety alignment. To address this problem, we propose Scalable Bilevel Data Selection (S-BDS) for safety-preserving medical SFT. By controlling selection-weight concentration, S-BDS enables reliable sample ranking over large candidate pools and identifies training examples that improve medical capability while controlling patient-facing risks. We combine clinician-validated safe data and selected data with unscreened medical instruction data for SFT, and jointly evaluate medical capability and patient-facing safety on HealthBench and PatientSafetyBench, respectively. Experiments across Llama and Qwen models show that S-BDS achieves the highest safe rates among all compared methods, outperforming the strongest baselines by 4.64 and 3.00 percentage points, respectively, while maintaining competitive medical capability. Further experiments show that varying the safety-oriented weight during SFT controls the trade-off between medical capability and patient-facing safety.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include SFT, HealthBench, Llama, Qwen.

Source attribution: curated from arXiv cs.AI (@Ruoxuan Shi, Sheng Yang, Zhengxing Su, Xiaoyang Hou, Yating Liu).