Effective Data Balancing and Fine-Tuning Techniques for Medical sLLMs in Resource-Constrained Domains

Abstract

Despite remarkable advances in medical large language models (LLMs), their deployment in real clinical settings remains impractical due to prohibitive computational requirements and privacy regulations that restrict cloud-based solutions. Small LLMs (sLLMs) offer a promising alternative for on-premise deployment, yet they require domain-specific fine-tuning that still exceeds the hardware capacity of most healthcare institutions. Furthermore, the impact of multilingual data composition on medical sLLM performance remains poorly understood. We present a resource-efficient fine-tuning pipeline that integrates Quantized Low-Rank Adaptation (QLoRA), Fully Sharded Data Parallelism (FSDP), and Sequence Packing, validated across two model scales: MedGemma 4B for efficiency analysis and LLaMA 3.3 70B for data balance experiments. Our approach achieves 58.3% reduction in video random access memory (VRAM) usage (from 48 GB to 20 GB) and 5× training speedup on MedGemma 4B using NVIDIA L40s GPUs. Critically, experiments on LLaMA 3.3 70B reveal that English-heavy data mixing (10:3 ratio) degrades Korean medical law performance by 1.23 percentage points while providing only marginal English gains (+1.49 pp), demonstrating catastrophic forgetting in multilingual medical fine-tuning. Our work provides three contributions: (1) a practical fine-tuning pipeline operable within 20 GB VRAM, (2) empirical evidence that data balance—not volume—determines multilingual medical QA performance, and (3) actionable guidelines for deploying medical sLLMs in non-English clinical environments.

Ähnliche Arbeiten