LeMaterial/Atompack

Atompack是Hugging Face Hub上用于分发原子级机器学习（atomistic machine learning）公共数据集的存储库，采用Atompack格式。该存储库是更广泛的LeMaterial项目的一部分，主要负责数据的分发和服务。数据集来源于多个上游公共数据集，包括LeMat-Bulk、MatPES、MP-ALOE、MPtrj和OMAT24。该存储库并非这些数据集的原始来源。Atompack设计用于处理原子级数据集从小型科学数据库到训练语料库的转变，支持只读mmap访问、完整分子记录的索引读取、本地文件和分片目录的支持，以及Hugging Face Hub路径的直接打开和下载帮助。

Atompack is the Hugging Face Hub repository for public atomistic ML datasets distributed in the Atompack format. This repository is part of the broader LeMaterial effort. Its role is distribution and serving: it exposes packaged dataset paths that can be opened directly with the `atompack` Python package. The data hosted here comes from upstream public datasets such as LeMat-Bulk, MatPES, MP-ALOE, MPtrj, and OMAT24. This repository is not the original source of those datasets. Atompack is designed for the point where atomistic datasets stop behaving like small scientific databases and start behaving like training corpora, providing read-only mmap-backed access, direct indexed reads of full molecule records, support for local files and shard directories, and direct open/download helpers for Hugging Face Hub paths.