DRD: Chinese Diplomatic Rhetoric Dataset for Supervised Fine-Tuning of Large Language Models

The dataset derived from the routine press conferences held by the spokespersons of China's Ministry of Foreign Affairs between 2000 and 2024. A total of 20,745 Q&A pairs were collected and curated, forming a comprehensive Chinese Diplomatic Rhetoric Dataset (DRD) intended for supervised fine-tuning of large language models. The aim is to provide a specialized dataset on diplomatic dialogue strategies to enhance existing Chinese language models, enabling them to more accurately comprehend and respond to diplomatic discourse within an international context. The paper introduces the N-Jaccard text similarity algorithm, which mitigates the sensitivity to text length and considers the order of words, thereby capturing semantic relationships and logical connections more effectively. Utilizing the GPT-3.5 Turbo large language model, core information was extracted from the data, converted into contextually appropriate topics, and refined to generate input prompts and output responses. Both machine and human reviewers audited and corrected the initial processed data to ensure the dataset's quality. The DRD dataset, based on extensive high-quality Q&A data, effectively reflects the conciseness, accuracy, and artistry of diplomatic language expression.

本数据集源自2000年至2024年间中国外交部发言人举行的例行记者会，共收集整理20745组问答对，构建了面向大语言模型（Large Language Model）监督微调的综合型中国外交话语数据集（Diplomatic Rhetoric Dataset，DRD）。本数据集旨在提供专门化的外交对话策略语料，以优化现有中文大语言模型，使其能够更精准地理解并回应国际语境下的外交话语。本文引入N-Jaccard文本相似度算法，该算法可降低对文本长度的敏感性并兼顾词序，从而更有效地捕捉语义关联与逻辑联系。研究依托GPT-3.5 Turbo大语言模型，从原始数据中提取核心信息，转换为契合语境的主题，并经精细化处理生成输入提示与输出应答。研究团队通过机器与人工双重审核对初始处理后的数据进行校正，以保障数据集的质量。本数据集基于大规模高质量问答语料，充分体现了外交语言表达的简洁性、准确性与艺术性。

搜集汇总

背景与挑战

背景概述

该数据集是一个专门用于大语言模型监督微调的中文外交修辞数据集，包含2000年至2024年中国外交部发言人例行记者会的20,745个问答对。它通过N-Jaccard算法和GPT-3.5 Turbo模型处理，并经过人工审核，旨在提升模型对外交话语的理解和响应能力，反映外交语言的简洁性、准确性和艺术性。