Skip to main content
Back to timeline
arXivSource publication:

Medical LLM fine-tuning no longer trades away patient safety: S-BDS lifts PatientSafetyBench safe rates by 4.64 points

Synopsis

Patient-facing medical LLMs usually gain domain knowledge through supervised fine-tuning on large medical instruction sets, but that process can erode their existing safety alignment. S-BDS adds entropy regularization to the upper-level objective of bilevel data selection, curbing excessive concentration of selection weights so that sample rankings stay reliable over large candidate pools, and uses clinician-validated safe data as the upper-level signal. On Llama-3.1-8B-Instruct and Qwen3-8B-Instruct, S-BDS achieves the highest PatientSafetyBench safe rates among compared methods, 4.64 and 3.00 percentage points above the strongest baseline SEAL, while keeping competitive HealthBench medical capability scores.

Interpretation

Medical instruction tuning can expand safety-oriented data without per-example clinician review: S-BDS uses a small set of clinician-confirmed safe examples as guidance to pick training samples from a large candidate pool that improve medical capability while controlling patient-facing risk. Earlier patient-safety evaluation was largely diagnostic, identifying unsafe behavior in trained models but offering no mechanism to preserve patient-facing safety when a general LLM undergoes medical supervised fine-tuning; the existing bilevel selection method SEAL also did not explicitly control how concentrated the selection weights became. On Llama-3.1-8B-Instruct and Qwen3-8B-Instruct, S-BDS reaches PatientSafetyBench safe rates of 69.31% and 85.62%, above the strongest baseline SEAL at 64.67% and 82.62%, a gain of 4.64 and 3.00 percentage points respectively.

The method casts data selection as a two-level problem: the lower level fine-tunes the model on candidate medical data, and the upper level rewards sample weights that preserve performance on patient-safe supervision. Solving this directly requires differentiating through the lower-level optimization, so S-BDS instead relaxes it with a lower-level optimality gap and adds entropy regularization to the upper-level objective, keeping the normalized selection-weight distribution from collapsing onto a few examples. In the Orca setting, the original SEAL selector gives the top 5% of Llama3.1 candidate examples 42.15% of total selector mass and the top 20% 64.16%; S-BDS reduces these to 13.48% and 36.61% and raises normalized effective sample size from 20.62% to 78.25%.

Final fine-tuning merges clinician-validated safe data with the selected examples into a safety-oriented dataset, then mixes it by weight with unscreened medical task data. Over-weighting safety data can cause over-refusal or capability loss, while task-only fine-tuning weakens existing safety alignment, so both kinds of supervision need to be retained. Under a fixed training budget, raising the safety-data weight from 0 to 0.5 improves both Health Score and PSB safe rate for both models; beyond 0.5 the safe rate keeps rising while Health Score declines.

Perspective

Experiments are limited to English, single-turn, patient-facing medical question answering, with fine-tuning data from HealthCareMagic-100k, base models Llama-3.1-8B-Instruct and Qwen3-8B-Instruct, and LoRA-based supervised fine-tuning; safety is evaluated on the 466 single-turn patient queries of PatientSafetyBench and capability on HealthBench.

The authors note that this controlled setting may not capture multilingual or multi-turn medical consultations, or broader clinical capabilities such as longitudinal care and clinical information extraction, and that reliance on a single instruction-data source may limit generalization; LLM judgments agreed with clinicians on 88.8% of 480 sampled responses overall, with Cohen's κ of 0.73, ranging from 82.5% to 97.5% and κ from 0.52 to 0.94 across model-method groups.