CoDI-Eval

CoDI-Eval是由中国科学技术大学开发的一个大型语言模型评估数据集，专注于多样化的指令遵循能力。该数据集包含10000条指令，涵盖情感、主题、长度、关键词和毒性避免等多个控制生成任务。创建过程中，采用了两步法：首先通过扩展过程增加指令数量，然后通过多样化过程增加指令表达的多样性。CoDI-Eval的应用领域包括评估和比较不同语言模型在遵循特定指令时的表现，旨在提高模型在复杂指令下的控制生成能力。

CoDI-Eval is a large language model evaluation dataset developed by the University of Science and Technology of China, focusing on diverse instruction-following capabilities. This dataset contains 10,000 instructions, covering multiple controlled generation tasks including sentiment control, theme control, length constraint, keyword constraint, and toxicity avoidance. It was built via a two-step workflow: first, expand the number of instructions through an expansion process, and then boost the diversity of instruction expressions through a diversification process. The application fields of CoDI-Eval include evaluating and comparing the performance of different large language models when complying with specific instructions, with the goal of improving the controlled generation abilities of models under complex instruction scenarios.