ChikungunyaQA: A High-Fidelity, Multi-Persona Clinical Question-Answering Dataset Derived from Brazilian Ministry of Health Guidelines

ChikungunyaQA is a high-fidelity, expert-validated question-answering (QA) dataset focused on the clinical management of Chikungunya virus disease in the Brazilian public health context. It contains 1,147 QA pairs entirely grounded in official clinical manuals published by the Brazilian Ministry of Health (Ministério da Saúde). The dataset was constructed using the Inductive Clinical Distillation framework, a pipeline that combines a 3-stage Advanced Retrieval-Augmented Generation (RAG) architecture—Hybrid Search (BM25 + Dense Vectors) with Neural Cross-Encoder Re-ranking—with an iterative Saturation Mode generation strategy. Each source document chunk is processed in up to six successive passes, with an Incremental Semantic Memory mechanism that prevents the generation of duplicate questions, ensuring exhaustive knowledge extraction. Every QA pair was generated by a Multi-Persona Architecture that explicitly tailors the language register and level of detail to three distinct target audiences: Médico / Physician — Technical clinical language suitable for healthcare professionals. Paciente / Patient — Plain-language explanations accessible to the general public. Cuidador / Caregiver — Action-oriented instructions for family members and informal caregivers. Quality assurance was performed through an automated G-Eval (LLM-as-a-Judge) pipeline using a double-judge setup (GPT-4o-mini-2024-07-18 as generator; Claude Sonnet as auditor), with adaptive rescue cycles for borderline cases. The resulting gold-standard dataset achieved a mean G-Eval score of 4.78 / 5.0. A subset of 126 QA pairs underwent additional human expert review, with annotators assigning independent quality scores (0–100) to measure inter-rater agreement (Spearman's ρ between G-Eval and human scores). Source documents (5 clinical guidelines) are included: Document Description Manejo_Chikungunya_2ed.md Brazilian MoH Clinical Management Manual for Chikungunya, 2nd edition (primary source) IXCHIQ (vacina chikungunya) ANVISA technical dossier for the IXCHIQ vaccine Nota Técnica nº 28…2023 Technical note on Chikungunya surveillance Nota Técnica nº 64/2023 Technical note on arbovirus clinical protocols Nota Técnica nº 9/2026 Updated technical note on Chikungunya management Clinical domains covered by the QA pairs include: diagnosis criteria, laboratory confirmation methods, clinical phases (acute, post-acute, chronic), risk groups (elderly, neonates, pregnant women), treatment protocols (analgesics, NSAIDs, corticosteroids, hydroxychloroquine, methotrexate), vaccination (IXCHIQ), neurological complications, differential diagnosis with Dengue and Zika, and epidemiological surveillance. This dataset is intended to support research in: clinical NLP for low-resource languages, biomedical QA benchmarking, RAG system evaluation, LLM fine-tuning for the Brazilian Portuguese medical domain, and health informatics applied to tropical infectious diseases. Upload Type / Files The following files should be uploaded to Zenodo: File Description ChikugunyaQA.jsonl Gold-standard QA dataset (1,147 pairs, JSONL format) ChikugunyaQA.csv Gold-standard QA dataset in tabular CSV format ChikugunyaQA_alpaca.jsonl Gold-standard QA dataset in Alpaca instruction format for fine-tuning Chikungunya_human_validation.csv Full human annotation file (114 pairs) containing the human validation scores ("notas")