Low-Resource-RAG-Assistant-for-Naval-Maintenance-datasets

This dataset contains five JSON collections supporting the paper "An Offline Low-Resource RAG Assistant for Naval Maintenance: Bridging Spanish Queries and English Technical Documentation." 1. dataset_completo.json — Complete synthetic question-answer-context triples (250 samples) generated from English naval maintenance manuals using the RAGalyst framework. This is the raw, unpartitioned corpus used to produce the training and validation splits. 2. train_dataset.json — Processed training split (80%) ready for embedding model fine-tuning. Formatted for sentence-transformers with anchor, positive, and negative pairs using MultipleNegativesRankingLoss. 3. val_dataset.json — Processed validation split (20%) used for early stopping during embedding model training. Follows the same format as train_dataset.json. 4. data_en.json — English evaluation set (244 question-answer-context triples) used as the source benchmark and for monolingual retrieval validation. 5. data_es.json — Spanish evaluation set (244 questions with translated answers) derived from data_en.json using Qwen3-235B-A22B. Used for cross-lingual retrieval evaluation (Phase 1) and end-to-end RAG ablation (Phase 2). All data are synthetic question-answer pairs derived from technical naval maintenance documentation. No full proprietary manuals are included. These datasets are derived from proprietary naval maintenance documentation owned by Cotecmar S.A. They are provided for academic research and reproducibility purposes only. Commercial use, redistribution, or modification without explicit written consent from Cotecmar is strictly prohibited. Access is granted on a case-by-case basis upon request.