Installation

pip install goldilocks-ml

That installs everything, including PyTorch, pymatgen and the rest of the scientific stack the real models need. It's enough to use a published model and to train with the built-in reference trainers or the real ones.

The is_magnetic classifier needs one more, manual step

is_magnetic reads a frozen mMACE backbone's embedding, which needs mace on top of the install above. There is no extra for it: the mace-torch this needs is a research collaborator's fork (CheukHinHoJerry/mace, not the upstream ACEsuit/mace package of the same name on PyPI), which has no PyPI release at all -- and a package published to PyPI cannot declare a direct git dependency in its own metadata regardless. Install it by hand:

pip install ase==3.28.0 e3nn==0.4.4 sphericart==1.0.9 sphericart-torch==1.0.9
pip install "mace-torch @ git+https://github.com/CheukHinHoJerry/mace.git@19cdf6692c48e068a24e06cfe1ffc670e8aea3dd"

See deposits/magnetism/is_magnetic/mace_mlp/VENDORING_TODO.md in the repository for the full story on that pinned commit.

Check it worked

goldilocks-ml --version
goldilocks-ml --help

You should see the two things this tool does: train and publish.

Working on Goldilocks ML itself

Clone the repository and let uv build the locked environment:

git clone https://github.com/stfc/goldilocks-ml.git
cd goldilocks-ml
uv sync --group dev
uv run pytest

Commands in these docs are written as uv run goldilocks-ml … because they run from a clone. With the package installed, drop the uv run.

Continue with Use a model or Train a model.