PrivateLabBench Adapters

Adapters connect domain-specific data or models to the common local evaluation workflow.

The simplest integration remains a local prediction table. Model adapters are useful when a reproducible built-in baseline or direct local model integration adds value.

Extension model

The generic ModelAdapter protocol no longer assumes a molecule dataset. Core extension interfaces live under privatelabbench.core.

Domain packages should depend on these interfaces rather than adding special cases to the central runner.

Built-in molecule adapters

Hashed Random Forest

model:
  adapter: hashed_random_forest
  fingerprint: hashed
  n_bits: 256
  n_estimators: 200

RDKit Morgan Random Forest

model:
  adapter: rdkit_random_forest
  fingerprint: rdkit_morgan
  n_bits: 2048
  radius: 2
  n_estimators: 200

Install with pip install -e '.[rdkit]'.

External predictions

plb eval-predictions predictions.csv --target target --prediction-column prediction

Third-party tasks

Packages can register tasks through the Python entry-point group privatelabbench.tasks.

Useful contributions include scientific foundation-model adapters, proteins, materials, microscopy, robotics/trajectory evaluation, and better molecular featurizers. Heavy dependencies should remain optional.