llylly001/InfiBench

InfiBench是一个用于评估代码大型语言模型（LLMs）问答能力的数据集。数据来源于StackExchange的公开存档，经过预处理和筛选，包含1,090,238个问题，最终筛选出13,854个问题作为初始种子集。数据集由五位领域专家进行标注，标注过程包括问题选择与类型标注、提示改写和正确性标准标注。数据偏见包括非标准评估、使用误解和潜在的数据污染。个人敏感信息在数据处理过程中被移除。

InfiBench is an evaluation dataset for the question-answering capabilities of code large language models. The dataset contains data downloaded and preprocessed from the StackExchange archive, specifically from StackOverflow posts formatted in Markdown text. The dataset selected questions that have at least three positively voted answers and an officially accepted answer, resulting in 13,854 questions out of 1,090,238. These questions were then randomly sampled and benchmarked by domain experts within the company. The dataset has potential data biases such as non-standard evaluation, usage misinterpretation, and potential data contamination. The dataset also pays special attention to the removal of personal sensitive information.

原始信息汇总

数据集许可证信息

• 许可证类型: CC-BY-SA-4.0
• 许可证说明: 该数据集遵循知识共享署名-相同方式共享4.0国际许可协议。