数据堂—10万对中文大模型通用领域复杂指令跟随SFT文本数据集

10万对中文复杂的prompt指令，字数在50~400字之间，每个prompt中的约束条件不低于3个，用于训练提升大模型的指令跟随性。类别覆盖生成类（写新闻稿、采访提纲、文案创作、文稿校对、中英作文、语法学习、研究报告、学习计划、诗歌创作、美食介绍、广告软文、销售话术、公文辅助写作、公文审核、政策文件问答等）、重写类（改写句子、文本纠错、句子合并、简化文案）、摘要类（内容摘要）、提取类（事件要素提取、观点提取、关键词提取、立场抽取、实体抽取）。所有prompt均为人工编写，满足多样性覆盖

100,000 pairs of complex Chinese prompt instructions, each containing between 50 and 400 Chinese characters, with no fewer than 3 constraint conditions per prompt. This dataset is designed to train large language models to improve their instruction-following capabilities. The covered task categories include generation tasks (e.g., news script writing, interview outline drafting, copy creation, manuscript proofreading, Chinese and English composition, grammar learning, research report writing, study plan formulation, poetry creation, food introduction, advertorial copywriting, sales pitch writing, official document auxiliary writing, official document review, policy document Q&A), rewriting tasks (e.g., sentence rewriting, text error correction, sentence merging, copy simplification), summarization tasks (content summarization), and extraction tasks (e.g., event element extraction, opinion extraction, keyword extraction, stance extraction, entity extraction). All prompts are manually developed to ensure comprehensive coverage of diverse task scenarios.

搜集汇总

背景与挑战

背景概述

该数据集是一个包含10万对中文复杂指令的SFT文本数据集，专门用于训练提升大模型在通用领域的指令跟随能力。每个指令字数在50-400字之间，且包含至少3个约束条件，覆盖生成、重写、摘要和提取等多种任务类别，所有指令均为人工编写以确保多样性和质量。