data-uniformity
收藏资源简介:
本文中描述的数据集名为 'data-uniformity',旨在研究数据均匀性对训练效率和性能的影响。数据集包含通过最大化数据点对之间的距离进行选择的数据,以探索数据均匀性如何加速训练并提高大型语言模型(LLMs)的性能。数据集由约翰霍普金斯大学应用数学与统计系和加州大学伯克利分校电子工程与计算机科学系的研究人员创建,并可在指定的GitHub链接中访问。
The dataset described in this paper is named 'data-uniformity', which aims to investigate the impact of data uniformity on training efficiency and performance. It consists of data selected by maximizing the pairwise distances between data points, to explore how data uniformity accelerates training and enhances the performance of Large Language Models (LLMs). This dataset was developed by researchers from the Department of Applied Mathematics and Statistics at Johns Hopkins University and the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley, and is accessible via the specified GitHub link.
数据集概述
数据集名称
Data Uniformity Improves Training Efficiency and More, with a Convergence Framework Beyond the NTK Regime
数据集用途
- 用于研究数据均匀性对训练效率的影响
- 提供超越NTK体系的收敛框架
数据集内容
- 包含TeaMs-RL数据集示例(full_dataset_9k.json)
- 支持LLaMA-7B模型的训练和转换
数据集使用方式
环境配置
- 需要安装Python 3.10环境
- 依赖库包括:
- PyTorch 1.12.0
- torchvision 0.13.0
- torchaudio 0.12.0
- CUDA Toolkit 11.3
训练步骤
- 下载数据集(如TeaMs-RL)
- 转换LLaMA检查点到HuggingFace格式
- 训练LLaMA-7B模型(支持单节点和多节点训练)
引用信息
bash @article{wang2025uniformity, title={Data Uniformity Improves Training Efficiency and More, with a Convergence Framework Beyond the NTK Regime}, author={Wang, Yuqing and Gu, Shangding}, journal={Github}, year={2025} }
致谢
- Llama-X
- Meta AI LLaMA
- Huggingface Transformers Llama
- Alpaca 和 Alpaca-LoRA




