Supporting data for expert-validated RAG dataset construction framework (naval case study N=44 + RAGBench evaluation package) — restricted

Restricted supporting data for the manuscript: "An End-to-End Framework for Expert-Validated Dataset Construction in Domain-Specific Retrieval-Augmented Generation Systems"(Journal of Intelligent Information Systems). This record contains two complementary packages: 1) Naval maintenance case-study dataset (RESTRICTED / confidential)- 44 expert-validated question–answer–context triples from routine naval technical maintenance consultations.- Breakdown: Caterpillar engine (24), desalination system (17), Mercury propulsion (3).- Fields: question, answer, document_name, contexts.- Spanish questions/answers; English technical-manual excerpts in contexts.- Does not include full manuals, usernames, roles, IP addresses, or raw chat logs. 2) Public benchmark reproducibility package (RAGBench evaluation artifacts)- File: ragbench_pruner_eval_reproducibility.zip- Shared RAGBench test split (N=480), evaluation reports for the proposed bilingual sentence pruner, cross-encoder reranker, and two LLM-as-judge baselines, paired significance outputs, and scripts to recompute the reported tests.- Derived from the public RAGBench benchmark; cite the original RAGBench paper when reusing. Access:The entire Zenodo record is currently RESTRICTED. Files are available on reasonable request to the corresponding author (Anderson Morillo, [e-mail]), subject to institutional approval for the naval case-study data. The naval triples must not be redistributed without written permission from Cotecmar / Universidad Tecnológica de Bolívar.