hseb-benchmark/msmarco

HSEB MSMARCO基准数据集是基于MSMARCO数据集构建的，包含三种不同维度的嵌入模型：384维的all-MiniLM-L6-v2，768维的intfloat/e5-base-v2和2560维的Qwen3-Embedding-4B。数据集分为三个规模：1K（1000个文档，10000个查询），100K（100000个文档，10000个查询）和1M（1000000个文档，10000个查询）。每个文档根据采样选择性有一个标签，选择性分为10%，90%和100%。对于每个查询和每个选择性级别，都有预先计算的确切k-NN搜索结果。

The HSEB MSMARCO benchmarking dataset is based on the MSMARCO dataset and includes three different embedding models: all-MiniLM-L6-v2 (384 dims), intfloat/e5-base-v2 (768 dims), and Qwen3-Embedding-4B (2560 dims). The dataset is split into three sizes: 1K (1000 documents, 10,000 queries), 100K (100,000 documents, 10,000 queries), and 1M (1,000,000 documents, 10,000 queries). Each document has a tag based on the sampled selectivity, with 10% selectivity for high, 90% for low, and 100% for no filters. For each query and selectivity level, there are precomputed exact k-NN search results for the top-100 documents.