Godseye1311/radial-hyena-networks

Radial Hyena — benchmark data, trained weights and training records

Companion files for Radial Hyena Networks for Property Prediction of Nanomaterials. The accompanying code downloads and verifies these files automatically.

Contents

Path	Contents
data/chili100k_benchmark.h5	the 2,975 CHILI-100K benchmark graphs: structures, labels, SAXS/XRD/xPDF curves and the train/validation/test split
checkpoints/<task>_seed<k>.pt	weights of the 18 reported models, task ∈ crystal_system, space_group, atom, saxs, xrd, xpdf; k ∈ 0, 1, 2
checkpoints/space_group_trainvocab_seed<k>.pt	space-group models with the train-split vocabulary (151 classes), used in the mechanistic analyses
records/	training records: per-epoch histories, split indices, configuration, software versions
MANIFEST.md5	MD5 checksum of every file

Verify a manual download from the repository root with md5sum -c MANIFEST.md5.

Benchmark subset

The official CHILI benchmark protocol: 425 graphs per crystal system (2,975 total), split 80/10/10 over individual graphs, stratified by space group, random_state=42 (2,379 / 298 / 298).

Data licence and attribution

The benchmark subset is derived from CHILI-100K and redistributed under its CC-BY-4.0 licence. Please cite the original dataset:

U. Friis-Jensen, F. L. Johansen, A. S. Anker, E. B. Dam, K. M. Ø. Jensen, R. Selvan. CHILI: Chemically-informed large-scale inorganic nanomaterials dataset for advancing graph machine learning. arXiv:2402.13221, 2024.