HELPSTEER

HELPSTEER是由NVIDIA创建的多属性帮助性数据集，包含37,120个样本，旨在通过详细标注如正确性、连贯性、复杂性和冗长性等属性，提升模型对帮助性响应的理解。数据集通过精心设计的流程收集，包括多样化的提示选择和人工标注，确保数据质量。该数据集适用于训练更有效的通用领域语言模型，特别是在提高模型响应的帮助性和准确性方面。

HELPSTEER is a multi-attribute helpfulness dataset created by NVIDIA, containing 37,120 samples. Its core objective is to enhance models' understanding of helpful responses through detailed annotations of attributes including correctness, coherence, complexity, and verbosity. The dataset is collected via a meticulously designed workflow, which encompasses diverse prompt selection and manual annotation to guarantee high data quality. This dataset is applicable for training more effective general-purpose language models, especially for improving the helpfulness and accuracy of the models' generated responses.