SuperNI

该数据集是一组通过不同语言模型生成的指令调优数据集，并在SuperNI测试集上进行评估。它包含了根据人工评估分类为好、坏和可能性的示例，规模大约有30,000个例子，其任务旨在进行指令调优和评估。

This dataset comprises approximately 30,000 instruction-tuning examples generated by various language models. These examples are classified into three categories: good, bad, and potential based on human evaluation, and the dataset is evaluated on the SuperNI test set. The tasks included in this dataset are designed for instruction tuning and model evaluation.

搜集汇总

背景与挑战

背景概述

SuperNI数据集是一个用于指令调优数据生成的数据集，由IBM发布，与EMNLP2023论文相关。它包含通过多个步骤生成的合成数据，包括指令生成、实例生成、过滤和输出集成，旨在提高输出质量。