Combinatorial optimization (CO) has been studied as a useful tool for modeling industrial problems, but it still remains a challenge in complex domains because of the NP-hardness. With recent advances in machine learning, the field of CO is shifting to the study of neural combinatorial optimization using a large amount of data, showing promising results in some CO problems. Rectilinear Steiner tree packing problem (RSTPP) is a well-known CO problem and is widely used in modeling wiring problem among components in a printed circuit board and an integrated circuit design. Despite the importance of its application, the lack of available data has restricted to fully leverage machine learning approaches. In this paper, we present ReSPack, a large-scale synthetic RSTPP data generator and a benchmark. ReSPack includes a source code for generating RSTPP instances of various types with different sizes, test instances generated for the benchmark evaluation, and implementations of several baseline algorithms.