TVA: A Triadic Gradient Preference Dataset for Value Alignment in Large Language Models

Each sample has four attributes: instruction, outputted as good, outputted as medium, and outputted as bad, corresponding to instruction, superior response, intermediate response, and inferior response, respectively. According to the rating verification of the large model, each triplet can be divided into three sets of binary preference data, totaling 23067 pieces of data, to facilitate model training.

All samples have been cleaned and desensitized. As the dataset contains negative and inferior samples used for violations, it is recommended to strictly follow the guidelines for academic research and safety alignment training only.