# LeMaterial/Atompack

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](https://lematerial.org) 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.

## Install

```bash
pip install atompack-db
```

Import the package as `atompack` after installation.

## Open a Dataset

```python
import atompack

db = atompack.hub.open(
    repo_id="LeMaterial/Atompack",
    path_in_repo="omat/train",
)
print(len(db))
print(db[0].energy)
db.close()
```

You can also download a local copy first:

```python
import atompack

local_path = atompack.hub.download(
    repo_id="LeMaterial/Atompack",
    path_in_repo="omat/train",
)
db = atompack.hub.open_path(local_path)
print(len(db))
db.close()
```

## Dataset Paths and Sources

The main public dataset paths currently exposed in this repository are:

- `lematbulk/pbe`, from [LeMat-Bulk](https://huggingface.co/datasets/LeMaterial/LeMat-Bulk). If you use this path, cite LeMat-Bulk.
- `matpes/pbe` and `matpes/r2scan`, from [MatPES](https://docs.materialsproject.org/collaborations/matpes). If you use these paths, cite MatPES.
- `mp_aloe`, from [MP-ALOE](https://www.nature.com/articles/s41524-025-01834-9). If you use this path, cite MP-ALOE.
- `mptrj`, from [MPtrj](https://docs.materialsproject.org/services/ml-and-ai-applications/mptrj). If you use this path, cite MPtrj.
- `omat/train` and `omat/val`, from [OMAT24](https://huggingface.co/datasets/facebook/OMAT24). If you use these paths, cite OMAT24.

These paths may be stored as one `.atp` file or as a shard directory.
`atompack.hub.open(...)` handles both through the same read-only API.

## Why Atompack

Atompack is designed for the point where atomistic datasets stop behaving like small scientific
databases and start behaving like training corpora: repeated random reads, multiprocessing
workers, large immutable snapshots, and regular export and publish steps.

It provides:

- read-only mmap-backed access for static datasets
- direct indexed reads of full molecule records
- support for local files and shard directories
- direct open/download helpers for Hugging Face Hub paths

## More

- Project repository: <https://github.com/LeMaterial/atompack>
- LeMaterial: <https://lematerial.org>