Layer Placement of LoRA Adapters in Qwen2.5 for English–Chinese Medical Question Answering

This study investigates the use of layer-wise Low-Rank Adaptation (LoRA) for mitigating cross-lingual factual hallucinations in medical question answering. The study evaluates the effect of adapter placement across different transformer layers of Qwen2.5-7B-Instruct, with a focus on English–Chinese medical question answering. Parameter-efficient fine-tuning is employed to compare layer-specific adaptation strategies under constrained computational resources. The evaluation considers factual accuracy and cross-lingual consistency to examine how the location of trainable LoRA adapters influences model performance. The results provide empirical insights into the relationship between layer placement and factual reliability in multilingual medical question answering and contribute to the broader study of parameter-efficient adaptation for large language models.