COIG

我们提出了中文开放教学通才 (COIG) 项目，以维护一套无害，有用和多样化的中文教学语料库。我们欢迎社区中的所有研究人员为语料库集做出贡献并与我们合作。我们仅发布了COIG的第一个芯片，以帮助中国llms在探索阶段的发展，并呼吁更多的研究人员加入我们的行列。我们引入了手动验证的翻译通用指令语料库，手动注释的考试指令语料库，人类价值对齐指令语料库，多轮反事实纠正聊天语料库和leetcode指令语料库。我们提供这些新的教学语料库，以协助社区对中文llm进行教学调整。这些指令语料库也是如何有效构建和扩展新中文指令语料库的模板工作流。

We introduce the Chinese Open Instruction Generalist (COIG) project, which aims to curate a harmless, useful, and diverse Chinese instruction corpus. We welcome all researchers in the community to contribute to this corpus collection and collaborate with us. We have only released the first installment of COIG to support the development of Chinese LLMs during their exploratory phase, and call on more researchers to join our project. We have compiled five types of corpora: manually verified general translation instruction corpus, manually annotated exam instruction corpus, human value-aligned instruction corpus, multi-turn counterfactual correction chat corpus, and LeetCode instruction corpus. We release these novel instruction corpora to assist the community in conducting instruction tuning for Chinese LLMs. These instruction corpora also serve as a template workflow for effectively constructing and expanding new Chinese instruction corpora.

搜集汇总

背景与挑战

背景概述

COIG是一个中文开放教学通才项目，致力于构建无害、有用且多样化的中文教学语料库，包含多种手动验证和注释的指令数据，以支持中文大语言模型的教学调整。该数据集由多所研究机构于2023年联合发布，采用Apache 2.0许可。