A value-aligned dataset (VAD) for supervised fine-tuning of large language models

Each JSON object has five attributes, namely instruction, input, output, source, and label, which correspond to the user's instruction, supplementary input for the task, LLM generated reply, data source, and data classification, respectively. Among them, the data source includes human glm4、alpaca3， Corresponding to manually constructed data, large model Chat-GLM4 generation, and Alpaca data translation.

The data classification label includes person social、nation， The categories of the data content are individual, social, and national.