CLI

Equitrain installs focused command-line entry points for each workflow.

Command	Purpose
equitrain	Train a model.
equitrain-preprocess	Convert input data to Equitrain HDF5 and compute statistics.
equitrain-evaluate	Evaluate a model on a test set.
equitrain-predict	Run prediction on an HDF5 dataset.
equitrain-export	Export a Torch model checkpoint, including supported fine-tuning adapters.
equitrain-inspect	Inspect model/checkpoint state.
equitrain-hdf5-info	Summarize Equitrain HDF5 files.
equitrain-hdf5-benchmark	Benchmark sequential HDF5 read throughput.

Every command supports -h/--help:

equitrain --help
equitrain-preprocess --help
equitrain-evaluate --help
equitrain-predict --help
equitrain-export --help

Use this page as an entry-point map. Detailed data conversion options are in Data and Preprocessing, and training/checkpoint controls are in Training Options.

Common Arguments

Training, evaluation, and prediction share these model arguments:

• --backend: torch by default, or jax.
• --model: Torch model file path or JAX bundle path.
• --model-wrapper: wrapper name such as mace, ani, orb, sevennet, or m3gnet.
• --r-max: optional cutoff override.
• --dtype: default numeric dtype.

Torch workflows usually use --batch-size. JAX train/evaluate workflows require --batch-max-edges for fixed-shape graph batches; JAX prediction may also use --batch-max-nodes. See Training Options.

Loss Arguments

The standard loss weights are:

• --energy-weight
• --forces-weight
• --stress-weight

The loss function is selected with --loss-type and can be overridden per quantity with --loss-type-energy, --loss-type-forces, and --loss-type-stress.

Torch also supports reaction-relative losses:

• --barrier-weight
• --reaction-energy-weight

These are described in Reaction-Relative Losses.

JAX Distributed Arguments

JAX training can initialize jax.distributed:

equitrain \
    --backend jax \
    --distributed \
    --launcher none \
    --process-count <global-processes> \
    --process-index <rank> \
    --coordinator-address <host:port> \
    ...

On a single node with multiple visible GPUs, the default launcher behavior can spawn local JAX processes automatically. For multi-node jobs, launch one Equitrain process per JAX process and pass explicit process metadata.