# Add a dataset

Make a new candidate pool available to ALF by subclassing {py:class}`BaseDataset <alf_core.dataset.base_dataset.BaseDataset>`. The dataset supplies the
candidate pool, its labels, train/validation splits, and a query interface, so it can serve both
the search space and (in offline mode) the {py:class}`Oracle <alf_core.oracle.oracle.Oracle>`.

**Base class:** {py:class}`BaseDataset <alf_core.dataset.base_dataset.BaseDataset>`. **Key methods:** `load_dataset`, `query`.

## Steps

1. Subclass {py:class}`BaseDataset <alf_core.dataset.base_dataset.BaseDataset>` under `tools/alf_tools/datasets/`.
2. Implement `load_dataset` (build the candidate pool and labels) and `query`.
3. Pick a compatible {term}`modality <Modality>` (e.g. protein `sequence` or SMILES) so models line up.
4. Add tests and run the build / pre-commit checks.

## Reference notebook

- [Datasets](https://github.com/instadeepai/alf/blob/main/tutorials/extending_base_classes/datasets.ipynb): implement a custom {py:class}`BaseDataset <alf_core.dataset.base_dataset.BaseDataset>`.