THU-KEG/RM-Bench

RM-Bench是一个评估语言模型奖励模型的基准数据集，包含了具有独特标识的样本，每个样本包括一个提示和多个回应（选择的和被拒绝的），回应分为简洁风格和详细风格（纯文本和Markdown格式）。数据集覆盖了聊天、代码、数学、安全拒绝和安全响应等不同领域。

RM-Bench is a benchmark dataset for evaluating the reward models of language models, which includes samples with unique identifiers, each containing a prompt and multiple responses (chosen and rejected) in both concise and detailed styles (plain text and Markdown). The dataset covers various domains such as chat, code, math, safety-refuse, and safety-response.