ZhiqiGao/Text2Opt-Bench
收藏资源简介:
Text2Opt-Bench是一个用于评估大型语言模型在操作研究优化问题(如线性规划、混合整数规划、非线性规划等)上表现的基准数据集,其特点是包含由求解器验证的真实解。该数据集旨在揭示当前大型语言模型在绑定(即将问题描述中的系数、索引和参数具体化)能力上的瓶颈。数据集包含模板生成的问题和非结构化问题,每个问题都有详细的JSON结构描述,包括元数据、实例数据、自然语言描述、参考解决方案等。此外,数据集还提供了正确性标准和评估提示,确保评估的准确性和一致性。
Text2Opt-Bench is a benchmark for evaluating LLMs on operations-research optimization problems (LP, MILP, MIQP, nonlinear) with solver-verified ground truth. It is designed to expose binding as the bottleneck for current LLMs, where binding refers to grounding every coefficient, index, and parameter in the concrete problem data. The dataset includes template-generated problems and unstructured problems, each with detailed JSON schema descriptions covering metadata, instance data, natural-language problem statements, reference solutions, and more. It also provides correctness criteria and evaluation prompts to ensure accurate and consistent assessments.





