surgeai/ComplexConstraints
收藏资源简介:
Complex Constraints Benchmark Set是一个用于评估语言模型遵循复杂、多约束指令能力的基准测试集。它包含75个项目(编号从CIF-001到CIF-075),每个项目包括一个真实提示和10到40个评估标准(总计1559个),这些标准描述了正确响应必须满足的条件。这些标准旨在用于基于量规的评分(人工或LLM作为评判者),而不是精确匹配。数据集以单个宽格式CSV文件组织,每行代表一个项目,包含benchmark_id、prompt、use_case、instruction_type、prompt_style等字段,以及最多40个评估标准列(criterion_{i},未使用的列为空)。
A benchmark for evaluating how well language models follow complex, multi-constraint instructions. It contains 75 items (CIF-001–CIF-075), each a realistic prompt paired with 10–40 evaluation criteria (1,559 total) describing what a correct response must satisfy. Criteria are meant for rubric-based grading (human or LLM-as-a-judge), not exact match. The dataset is structured as a single wide-format CSV file with one item per row, including fields such as benchmark_id, prompt, use_case, instruction_type, prompt_style, and up to 40 criterion columns (criterion_{i}, with unused columns left empty).




