Cosmoscd/AccelEval

ORBench测试数据集是用于评估大型语言模型在CPU到CUDA代码加速任务上的基准测试数据。该数据集包含三种不同规模的测试数据：小规模（约100 MB，用于冒烟测试，包含43个任务）、中规模（约3.5 GB，用于主排行榜）和大规模（约8 GB，用于压力测试）。每个任务数据包括二进制输入张量（ORBench v2格式）、CPU基线的参考输出、CPU基线运行时间以及适用的查询请求文件。数据集旨在通过模拟实际代码加速场景，测试LLMs在生成高效CUDA代码方面的性能。

ORBench Test Data is a benchmark dataset for evaluating Large Language Models (LLMs) on CPU-to-CUDA code acceleration tasks. It includes test data in three sizes: small (~100 MB, for smoke tests with 43 tasks), medium (~3.5 GB, for the main leaderboard), and large (~8 GB, for stress testing). Each task contains binary input tensors (in ORBench v2 format), expected output from a CPU baseline, CPU baseline wall time, and per-call queries if applicable. The dataset is designed to assess the performance of LLMs in generating efficient CUDA code for acceleration scenarios.