MIXTURE
收藏资源简介:
MIXTURE是一个基于维基百科的指令蒸馏数据集,包含约144,884个样本,跨越五个任务类型,提供从多个低质量输入到单个高质量输出的层次映射。数据集创建过程包括从维基百科收集原始数据,生成高、低质量数据对,并通过LLM评分、多变种退化、跨主题融合和噪声注入等方式构建数据集。该数据集旨在解决低质量数据在指令微调中的有效利用问题,通过蒸馏和增强低质量数据,提高指令微调LLMs的效率和性能。
MIXTURE is a Wikipedia-based instruction distillation dataset comprising approximately 144,884 samples across five task types, which provides a hierarchical mapping from multiple low-quality inputs to a single high-quality output. The dataset construction process includes collecting raw data from Wikipedia, generating high-quality and low-quality data pairs, and building the dataset via methods such as LLM-based scoring, multi-variant degradation, cross-topic fusion, and noise injection. This dataset aims to solve the problem of effective utilization of low-quality data in instruction fine-tuning, and improves the efficiency and performance of instruction-fine-tuned LLMs by distilling and augmenting low-quality data.
数据集概述
基本信息
- 数据集名称: english_wikipedia
- 存储位置: https://huggingface.co/datasets/lucadiliello/english_wikipedia
- 数据格式: 结构化文本
数据特征
- 字段构成:
- filename: 文件名(字符串类型)
- maintext: 正文内容(字符串类型)
- source_domain: 来源域名(字符串类型)
- title: 标题(字符串类型)
- url: 网页地址(字符串类型)
数据规模
- 训练集:
- 样本数量: 4,184,712条
- 数据大小: 10,569,005,563字节
- 下载大小: 6,144,953,788字节
- 总数据集大小: 10,569,005,563字节
数据来源
- 内容类型: 英文维基百科文本数据
- 结构特征: 包含网页元数据和正文内容




