DS2-INSTRUCT
收藏资源简介:
DS2-INSTRUCT是由西北大学团队开发的领域特定指令合成框架,旨在解决大语言模型在专业领域适应性问题。该数据集通过零样本方法自动生成,覆盖数学、金融、生物医学等7个专业领域,采用任务导向关键词生成和布鲁姆分类法的认知层级设计确保数据多样性。其创新性在于无需人工标注或领域语料库,通过双向关键词扩展和检索增强技术构建知识体系,并利用自一致性验证保障数据质量。该数据集专为大语言模型在专业领域的指令微调设计,可显著提升模型在领域术语和推理模式上的表现。
DS2-INSTRUCT is a domain-specific instruction synthesis framework developed by the team from Northwest University, aiming to address the domain adaptation issues of large language models (LLMs). This dataset is automatically generated via zero-shot methods, covering 7 professional domains including mathematics, finance, biomedicine and other fields. It adopts task-oriented keyword generation and cognitive hierarchy design based on Bloom's Taxonomy to ensure data diversity. Its innovation lies in constructing a knowledge system through bidirectional keyword expansion and retrieval-augmented technology without requiring manual annotation or domain corpora, and guaranteeing data quality via self-consistency verification. This dataset is specifically designed for the instruction tuning of large language models in professional domains, and can significantly enhance the model's performance in domain terminology and reasoning patterns.
DS²-INSTRUCT 数据集概述
数据集简介
DS²-INSTRUCT 是一个用于大型语言模型指令微调的领域特定数据合成框架。该框架采用零样本方法,无需人工监督即可生成领域特定的指令微调数据集。
核心特性
- 零样本框架:无需人工监督。
- 领域特定:专注于生成特定领域的数据。
- 用途:用于大型语言模型的指令微调。
支持的任务与领域
| 任务 | 领域 |
|---|---|
cfa |
金融 |
gsm8k |
小学数学 |
math |
竞赛数学 |
pubmedqa |
生物医学 |
logiqa |
逻辑推理 |
gpqa |
研究生级科学 |
medqa |
医学 |
输出数据
最终输出是一个用于监督微调的数据集,保存路径为:output/{task}_sft_dataset.json

- 1DS$^2$-Instruct: Domain-Specific Data Synthesis for Large Language Models Instruction Tuning西北大学 · 2026年



