Dataset Card: T1x-IRC-8K

• Repository: huggingface.co/datasets/mathboylinlin/T1x-IRC-8K
• Status: repository created; data upload pending until manuscript acceptance.

Summary

T1x-IRC-8K is a dataset of intrinsic reaction coordinate (IRC) trajectories used to train and evaluate the MARC-TS transition-state prediction pipeline.

reactions              8,209
IRC geometries         1,088,725

Locked split (to be released with the dataset)

train         7,409
validation      390
test            410

OOF fold sizes
1,482 / 1,482 / 1,482 / 1,482 / 1,481

The split is stored as frozen dataset indices; reaction IDs are derived from the canonical HDF5 reaction_ids array at load time. Two files will accompany the HDF5 release:

t1x_irc8k_split_v1.0.json        frozen train/validation/test indices
t1x_irc8k_train_folds_v1.0.csv   frozen 5-fold OOF assignment for training indices

Format (planned)

The dataset will be released as a single HDF5 file containing, per reaction, the IRC trajectory geometries, atomic numbers/types, and the mass-weighted reaction coordinate used as the MARC-Field query variable. Exact field-level documentation will accompany the HuggingFace upload.

Release timeline

The dataset will be uploaded to the HuggingFace repository above upon acceptance of the associated manuscript, together with the code release described in CODE_AVAILABILITY.md. This card will be updated with final field-level documentation, licensing terms, and citation information at that time.

Intended use

The dataset is intended for research on machine-learned reaction path and transition-state prediction methods. It is not intended for production chemical safety or regulatory use.