README.md

---
license: other
task_categories:
- text-generation
- tabular-regression
tags:
- materials-science
- crystal-structure
- materials-project
- parquet
pretty_name: CrystalReasoner MP Data
---

# CrystalReasoner-MP

This dataset contains the Materials Project-derived reference table used by CrystalReasoner for crystal-structure generation, property-conditioned prompts, and evaluation. The file is stored as a Parquet export of the project shelve database so it can be downloaded and reconstructed locally.

## Files

- `MP_shelve.parquet`: Parquet export of the reference data table.

## Dataset Fields

The current Parquet file contains 200290 rows. These columns are from Materials Project:

```text
material_id
structure
energy_above_hull
formation_energy_per_atom
is_stable
band_gap
is_metal
efermi
homogeneous_poisson
description
formula_pretty
```

and these columns are from our MLIP:

```text
bulk_modulus
shear_modulus
thermal_expansion_300k
```

The `structure` column is serialized in the project simple-crystal text format. Some property columns may contain JSON-serialized values or missing values, depending on data availability.

## Usage

Download the Parquet file and place it at the expected project path:

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="CrystalReasoner/CrystalReasoner-MP",
    repo_type="dataset",
    filename="MP_shelve.parquet",
    local_dir="assets/MP",
)
print(path)
```

Then recreate the shelve database used by the training and evaluation code:

```bash
python scripts/parquet_to_shelve.py
```

The reconstructed `assets/MP/MP_shelve` database is used by the training datasets, generation code, and ground-truth evaluation metrics.

## License

This dataset is derived from the Materials Project database.

Please refer to the Materials Project Terms of Use:
https://next-gen.materialsproject.org/about/terms