Balancing Binary Biomedical QA: Parameter-Efficient Instruction Tuning on BioASQ and PubMedQA
Large language models (LLMs) are prone to class imbalance bias, often overpredicting "Yes" or "No" due to skewed data arrangements, endangering in safety-critical biomedical QA tasks. In clinical contexts, such bias is particularly troublesome in clinical applications, as it may misinform medical decisions by incorrectly classifying negative cases as positive. In this work, We propose a parameter-efficient instruction-tuning framework for biomedical Yes/No QA, employing 4-bit quantized LoRA adapters to fine-tune three key models - Phi-3 Mini Instruct, Qwen2.5-1.5B, and BioGPT.We use two separate biomedical QA benchmarks to evaluate our models: We utilize two benchmark datasets: BioASQ Yes/No, which provides finely crafted QA pairs with brief explanations, and PubMedQA Yes/No, built from questions constructed over full research abstracts. To ensure fairness and consistency, we use set segmented splits and a single prompt template for every model. Evaluation is conducted using four metrics: Accuracy, Balanced Accuracy, Cohen’s κ, and Matthews Correlation Coefficient (MCC). phi3_mini_instruct performs best on PubMedQA (89% accuracy, MCC = 0.77), whereas Qwen2.5-1.5B provides the most consistent performance on BioASQ (72% accuracy, MCC = 0.08). In contrast, BioGPT shows bias and underachieves the general-purpose models, even after being pretrained on biomedical text.
Paper
Full text
Balancing Binary Biomedical QA: Parameter-Efficient Instruction Tuning on BioASQ and PubMedQA
Semantic Scholar · 2026
Abstract
Large language models (LLMs) are prone to class imbalance bias, often overpredicting "Yes" or "No" due to skewed data arrangements, endangering in safety-critical biomedical QA tasks. In clinical contexts, such bias is particularly troublesome in clinical applications, as it may misinform medical decisions by incorrectly classifying negative cases as positive. In this work, We propose a parameter-efficient instruction-tuning framework for biomedical Yes/No QA, employing 4-bit quantized LoRA adapters to fine-tune three key models - Phi-3 Mini Instruct, Qwen2.5-1.5B, and BioGPT.We use two separate biomedical QA benchmarks to evaluate our models: We utilize two benchmark datasets: BioASQ Yes/No, which provides finely crafted QA pairs with brief explanations, and PubMedQA Yes/No, built from questions constructed over full research abstracts. To ensure fairness and consistency, we use set segmented splits and a single prompt template for every model. Evaluation is conducted using four metrics: Accuracy, Balanced Accuracy, Cohen’s κ, and Matthews Correlation Coefficient (MCC). phi3_mini_instruct performs best on PubMedQA (89% accuracy, MCC = 0.77), whereas Qwen2.5-1.5B provides the most consistent performance on BioASQ (72% accuracy, MCC = 0.08). In contrast, BioGPT shows bias and underachieves the general-purpose models, even after being pretrained on biomedical text.