Taywon/HH_length_biased_15k

该数据集名为HH_length_biased_15k，是Anthropic/hh-rlhf数据集的一个子集，用于研究人类反馈的影响。数据集包含15k个随机样本，其中976个样本被翻转以偏向长响应，以在偏好数据集中引入长度偏差。数据集分为训练集和测试集，训练集包含15k个样本，测试集包含6k个样本。每个数据实例包含chosen（选择的响应）、rejected（拒绝的响应）和flipped（是否翻转）三个字段。

This dataset is a subset of Anthropic/hh-rlhf, used for the paper Understanding impacts of human feedback via influence functions. It contains 15k samples, with 976 samples flipped towards lengthy responses to induce length bias. The dataset is split into a training set containing flipped samples and a test set containing unflipped samples. The features include chosen (the chosen response), rejected (the rejected response), and flipped (an indicator of whether the response was flipped).