totally-not-an-llm/EverythingLM-data-V2

EverythingLM V2数据集是一个多样化的指令数据集，包含1000个人类与助手的对话。这些对话集是使用evol-instruct和Orca的原则生成的，涵盖了广泛的主题和互动。与V1版本相比，V2版本的所有数据均由GPT4生成，采用了更高质量的生成流程，包括更人性化的种子提示、修复了脚本中的错误、增加了创意写作的多样性等。数据集还包括了多种指令类别，如推理、创意写作、通用知识等，并且已经过滤了OpenAI的对齐内容。数据集的特点包括长且详细的输出、人性化的创造力、链式推理以及复杂和具有挑战性的任务。未来计划包括训练Llama 7b和13b模型，以及生成包含更多类别和GPT-4的V2版本数据集。

The EverythingLM V2 dataset is a diverse instruction dataset containing 1,000 human-assistant dialogue sessions. These dialogues were developed based on the principles of evol-instruct and Orca, spanning a broad spectrum of topics and interaction patterns. Compared to its V1 iteration, all data in the V2 version is generated using GPT-4, with an upgraded generation pipeline that incorporates more human-like seed prompts, fixes script errors, and boosts diversity in creative writing, among other quality enhancements. The dataset covers multiple instruction categories including reasoning, creative writing, general knowledge, and more, and has been filtered to exclude OpenAI-aligned content. Key features of the dataset include long and detailed outputs, human-like creativity, chain-of-thought reasoning, as well as complex and challenging tasks. Future plans include training Llama 7b and 13b models, as well as generating an enhanced V2 dataset with additional categories and GPT-4 generated content.

原始信息汇总

数据集概述

数据集名称

• EverythingLM V2

数据集类型

• 多样化的指令数据集

数据集规模

• 包含1000个人类辅助对话

数据集生成方法

• 结合了evol-instruct和Orca原则

数据集内容

• 涵盖广泛的主题和交互