On October 24, 2026, a study titled “IndoNutriQA: An Indonesian Clinical Nutrition Question Answering Benchmark” was officially published on the arXiv platform (ID 2610.31256v1), introducing for the first time IndoNutriQA—a question-answering benchmark specifically designed for the clinical nutrition domain in Indonesian.
This benchmark consists of question–answer pairs manually curated by a team of 911 experts, all drawn from real clinical nutrition reports. Each pair is anchored in clinically relevant context and accompanied by structured metadata, including anthropometric assessments, medical diagnoses, document sources, and text chunking information. The question types strictly align with the Nutrition Care Process (NCP), covering eight categories of PES diagnostic statements: Problem, Etiology, Signs and Symptoms, and others.
All questions were answered by one of three registered dietitians, while the original reports were written by nutrition graduate interns; the answers are presented as free-form clinical narrative text, rather than simple extractive snippets. Expert validation studies indicate that 6%–10% of the reference answers are subject to debate. This dataset supports generative, extractive, and retrieval-augmented question-answering tasks, and is suitable for zero-shot and few-shot evaluations of large language models.
The paper reports fine-tuned baseline results based on models such as IndoBERT variants, mT5, and RAG, employing 5-fold cross-validation, learning curve analysis, and Bootstrap 95% confidence intervals; it also conducts zero-shot and five-shot evaluations on eight contemporary large language models. The study further calibrates the LLM-as-judge evaluation framework, finding that it approaches human expert levels in factuality and completeness, but is unreliable for clinical safety judgments; manual error analysis further reveals that this framework can overlook unsupported content that experts would identify.
IndoNutriQA has been publicly released as a pure evaluation resource (not a large-scale training corpus), with the aim of advancing research and development in Indonesian-language clinical natural language processing and nutrition decision-support systems. The dataset’s sources and methodological details are based entirely on the original arXiv paper, with no additional external information.