Datasets¶

The quantum-mechanical datasets used to train and test SO3LR Tauri are on Zenodo:

The data behind the legacy so3lr model and the JACS paper are at https://doi.org/10.5281/zenodo.14779793.

Both are a starting point for fine-tuning or training from scratch.

The bundled so3lr-tauri-s/m/l checkpoints record which data trained them. In each model directory under so3lr/models/:

hyperparameters.yaml (and .json): the training mixture under data.datasets, with each source's path on the training cluster and its sampling weight.
fine_tuning.json: the run the checkpoint was refined from.
data_splits.json: which loader built each split.

Conventions¶

The original so3lr model was not trained on energies, so with it only energy differences between structures of the same composition are meaningful.

Besides energies and forces, datasets can include these optional labels:

dipole_vec: molecular dipole vectors
hirshfeld_ratios / a0_ratios / hirsh_ratios: volume or polarizability ratios, mapped to the key the selected model uses
c6_ratios: per-atom C6 coefficient ratios

Missing labels give NaN metrics in evaluation and are skipped in training.

Structures with elements a model does not cover are rejected with an error (see Models).

Small inputs for the example notebooks ship with the repository under examples/data/.