self_instruct

Self Instruct旨在提升语言模型遵循自然语言指令的能力，通过模型自身生成的数据来创建大型指令数据集，从而在无需大量人工标注的情况下改进语言模型的指令执行能力。它包含通过OpenAI的`davinci`引擎生成的52k条指令和82K条实例输入输出，可用于对语言模型进行指令微调。此外，还包含252个专家编写的、以用户为导向的任务及其指令，以及从P3和Super Natural Instructions数据集中抽取的50k个示例，用于实现公开数据集上的对比。数据集以英文呈现，并提供prompt和completion等字段，支持指令训练等任务。该数据集采用Apache-2.0许可协议。Self-Instruct aims to enhance the instruction-following capabilities of language models. It constructs large-scale instruction datasets by generating training data using the model itself, thereby improving their instruction-following performance without requiring extensive manual annotation. It includes 52k instruction samples and 82k input-output instance pairs generated by OpenAI's `davinci` engine, which are suitable for instruction fine-tuning of language models. Additionally, it contains 252 user-oriented tasks and their corresponding expert-authored instructions, as well as 50k examples extracted from the P3 and Super Natural Instructions datasets for conducting comparative evaluations on public datasets. The dataset is presented in English, with fields such as `prompt` and `completion` to support tasks including instruction training. This dataset is licensed under the Apache-2.0 license.

搜集汇总

背景与挑战

背景概述

Self-Instruct是一个用于提升语言模型指令遵循能力的英文数据集，通过模型自生成方式构建，包含52k条指令和82K条实例输入输出，以及专家编写的252个用户导向任务和50k个公开数据集示例。该数据集支持指令微调，采用Apache-2.0许可，旨在减少人工标注依赖，但需注意数据可能包含模型生成的错误或偏差。