TFMat reproducibility assets

This dataset accompanies TFMat: Text-Conditioned Flow Matching for Crystal Structure Generation. It contains the data and selected final artifacts needed to reproduce the reported TFMat experiments.

Contents

• data/: CrystalFlow benchmark data used by the baseline experiments.
• data_text/: TFMat text-annotated benchmark splits and precomputed MatSciBERT embeddings.
• hydra_jobs/singlerun/: final checkpoints, Hydra configurations, scalers, evaluation metrics, and selected generated samples for the paper's CSP and DNG experiments.

The release includes final checkpoints for Carbon-24, Perov-5, and MP-20 CSP baseline/text models, and MP-20 DNG baseline/text models. It intentionally excludes periodic intermediate checkpoints and transient log files.

Usage

Clone or download this dataset next to the TFMat code repository so that the data/, data_text/, and hydra_jobs/ paths match the repository layout. Then follow the commands in the code repository README.

The original benchmark data are derived from the public CDVAE/CrystalFlow benchmark resources. Please respect the licenses and terms of the upstream datasets when reusing them.