Factorial results for an audit of the character-level evaluation paradigm for dense bi-encoder retrievers

This dataset accompanies an empirical audit of the character-level evaluation paradigm for dense bi-encoder retrievers. The audit is a factorial of six dense bi-encoder embedders, six chunking algorithms, and three chunk-size regimes on thirteen character-level evaluation resources, yielding 1,404 cells over 15,174 unique queries and approximately 8.2 million row-level observations at five retrieval depths. The deposit contains the master results parquet, the thirteen evaluation corpora, and the derived analysis tables used in the paper.本数据集配套针对密集双编码器检索器（dense bi-encoder retrievers）的字符级评估范式实证评测研究。本次评测采用析因实验设计，涵盖6款密集双编码器嵌入器、6种分块算法与3种分块尺寸方案，基于13个字符级评测资源开展实验，最终生成1404组实验单元格，依托15174个唯一查询语句，在5种检索深度下共计得到约820万条行级观测数据。本数据存储库包含论文所用的主结果Parquet文件、13个评测语料库以及衍生分析表格。