# MAD (Massive Atomic Diversity) Training Split, ColabFit Conversion

## Facts

- Provider: Arslan Mazitov, Sofiia Chorna, Guillaume Fraux, Marnik Bercx, Giovanni Pizzi, Sandip De, Michele Ceriotti (EPFL); ColabFit conversion
- License: CC-BY-4.0 (https://archive.materialscloud.org/records/xdsbt-a3r17)
- Size: 4 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/colabfit/Massive_Atomic_Diversity_MAD_train
- Access: public, readable without an account
- DOI: https://doi.org/10.60732/f5b6ea1b
- Category: Chemistry & Materials
- Source: https://doi.org/10.48550/arXiv.2506.19674
- Page: https://scigantic.com/archive/ce88b2e1-38d4-57f8-a66a-00569f76d961
- Keywords: Chemistry & Materials, interatomic potentials, Parquet, Hugging Face

## About

The training split of MAD, the Massive Atomic Diversity dataset from the COSMO lab at EPFL (Mazitov, Chorna, Fraux, Bercx, Pizzi, De, Ceriotti). Starting from small sets of stable inorganic and molecular structures (Materials Cloud 3D and 2D databases, molecular crystals and fragments), the authors aggressively distort configurations with little regard for stability and compute them with consistent electronic structure settings, aiming for coverage of chemical and configurational space rather than accuracy on any one system. The creators report that, with fewer than 100,000 structures, MAD trains universal interatomic potentials competitive with models trained on far larger sets. Labels are energy, forces and stress. ColabFit's card counts 76,482 configurations; the repository is about 0.1 GB (measured 2026-11-14). This is the ColabFit Exchange conversion: configurations are Parquet files under co/ (structure, positions, cell, computed properties and metadata per row) with dataset-level information in ds.parquet; read them with pyarrow, pandas, polars or Hugging Face datasets, following the ColabFit configuration schema at materials.colabfit.org/docs. The Materials Cloud Archive record is CC-BY-4.0.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("colabfit/Massive_Atomic_Diversity_MAD_train")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `co/co_0.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/ce88b2e1-38d4-57f8-a66a-00569f76d961
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/ce88b2e1-38d4-57f8-a66a-00569f76d961
- Over MCP: `get_schema_card` with id `ce88b2e1-38d4-57f8-a66a-00569f76d961` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Chemistry & Materials datasets](https://scigantic.com/topics/chemistry-and-materials)
- [Interatomic potentials datasets](https://scigantic.com/topics/interatomic-potentials)
