GloriaGeng/REALM-Bench

该数据集是REALM-Bench中的作业车间调度（JSSP）基准，包含418个实例，用于评估大型语言模型和多智能体系统在现实世界动态规划和调度任务中的性能。数据集分为四个层级：J1（经典静态JSSP基准，最小化完工时间）、J2（J1实例加上动态干扰，如机器故障和停电，测试反应性重新规划）、J3（大规模静态实例，200个作业×50台机器）和J4（J3规模加上多个同时干扰，最高难度层级）。数据以统一JSON格式组织，每个实例包括作业数量、机器数量、作业操作序列（含机器和处理时间）以及干扰事件（如果适用）。数据集来源于标准运筹学基准（如DMU、TA、ABZ、SWV、YN），并包含清理和规范化处理。This dataset is the Job Shop Scheduling (JSSP) benchmark from REALM-Bench, comprising 418 instances designed for evaluating large language models and multi-agent systems on real-world, dynamic planning and scheduling tasks. It is organized into four tiers: J1 (classic static JSSP benchmarks minimizing makespan), J2 (J1 instances with dynamic disruptions such as machine breakdowns and power outages, testing reactive replanning), J3 (large-scale static instances with 200 jobs × 50 machines), and J4 (J3 scale with multiple simultaneous disruptions, representing the highest difficulty tier). Data is structured in a unified JSON format, with each instance including the number of jobs, number of machines, a sequence of job operations (with machine and processing time), and disruption events (if applicable). The dataset is sourced from standard operations research benchmarks (e.g., DMU, TA, ABZ, SWV, YN) and has been cleaned and normalized.