NamBoQA: a Vietnamese question answering dataset for tourism and cultural heritage of Southern Vietnam (data, code and results)

Data, code and results accompanying the manuscript "A Vietnamese question answering dataset and retrieval-augmented evaluation framework for tourism and cultural heritage of Southern Vietnam". NamBoQA is a Vietnamese question answering dataset on tourism and cultural heritage sites in eight provinces of Southern Vietnam (Nam Bộ). The retrieval corpus indexes 2,638 passages of 1,000 places. The gold set has 2,445 human-reviewed question–answer pairs on 318 places in four types (factoid 1,491; unanswerable 686; multi-hop 236; comparison 32). A further 6,916 human-reviewed silver pairs (1,466 newly generated on non-gold places and 5,450 paraphrases) are provided for training only, together with 8,839 pairs accepted by an LLM judge alone. All pairs were generated by a large language model behind an automatic citation-check gate and then reviewed by people; per-reviewer labels for every candidate (including rejected ones), adjudication outcomes and agreement statistics (Gwet's AC1 0.79–0.82 between reviewer pairs) are included. The archive also contains the data pipeline, the experiment code (BM25, Vietnamese bi-encoder zero-shot and fine-tuned, hybrid retrieval, extractive XLM-R reader, generative Gemini reader), the place-level train/dev/test split stratified by province, and the per-question outputs of all 12 configurations reported in the paper (three seeds for the main configurations, bootstrap tests). Main results on 695 test questions: hybrid Recall@5 0.881 ± 0.003; extractive F1 0.901 with gold context and 0.679 with top-5 retrieved context; generative F1 0.814 on factoid questions with retrieved context. Corpus licensing. Only the 1,119 passages derived entirely from Vietnamese Wikipedia (CC BY-SA 4.0) are redistributed verbatim. Text from provincial tourism portals and commercial travel websites is withheld; those passages keep their identifiers and metadata, and source URLs with access dates are listed per place. Questions, answers and short evidence quotes are released for all pairs. Data: CC BY-SA 4.0. Code: MIT (see LICENSE-CODE in the archive).