Mvanypersele/luciole_rag_benchmark

该数据集名为RAG citation benchmarks，是一个用于引用和检索基础问答评估的提示无关RAG基准导出。它包含多个配置，每个配置对应一个源数据集，如hotpotqa、hotpotqa_fr、newsquadfr、piaf、squad2_fr_pragnakalp和tatqa。每个数据行包括以下字段：id（稳定示例标识符）、query（用户问题）、retrieved_documents（检索到的上下文块列表）、titles（与检索文档对齐的块标题列表）、supporting_index（黄金或支持块的零基索引列表）和answer（预期答案；如果答案和支持都为空，则表示该行不可回答）。数据集提供了训练、验证和测试分割，具体行数和文件大小在README中详细列出。该数据集旨在支持RAG系统的评估，特别是在引用和检索基础QA任务中。

The dataset is titled RAG citation benchmarks and is described as prompt-agnostic RAG benchmark exports for citation and retrieval-grounded QA evaluation. It includes multiple configurations corresponding to source datasets: hotpotqa, hotpotqa_fr, newsquadfr, piaf, squad2_fr_pragnakalp, and tatqa. Each row contains the following fields: id (stable example id), query (user question), retrieved_documents (retrieved context chunks), titles (chunk titles aligned with retrieved_documents), supporting_index (zero-based indices of gold/supporting chunks), and answer (expected answer; an empty answer with empty supports represents an unanswerable row). The dataset provides splits for training, validation, and testing, with detailed row counts and file sizes listed in the README. It is designed for evaluating RAG systems, particularly in citation and retrieval-grounded QA tasks.