Contribute

Adding new targets

A target should be structured as a (sub-)Python package as follows:

target_package_name/
├── __init__.py
│   contains:
│       from .impl_module import TargetClassName
│       __all__ = ["TargetClassName"]
└── impl_module.py
    contains the class implementation TargetClassName

Templates to create targets can be found under:

https://github.com/ChristophervonKlitzing/boltzkit/tree/main/templates/targets

A target is a composition of capability providers that define how it behaves, for example whether it can evaluate densities or provide datasets. The minimal requirement is to inherit from BaseTarget.

Density providers

Define how a target evaluates probabilities and scores.

• NumpyDensityProvider NumPy-based implementation of log_prob and score functions.

• DispatchedDensityProvider Multi-backend implementation supporting NumPy, PyTorch, and JAX.

Dataset providers

Define how a target provides or constructs datasets.

• ExternalDatasetProvider Uses a manually specified dataloader such as CachedRepoDatasetLoader to load datasets from external sources like Hugging Face.

• ProceduralDatasetProvider Generates datasets procedurally (e.g. Gaussian mixture models).

Adding evaluation nodes