# Switch offline to online

The loop is identical in both modes; what changes is **where the labels come from**, which is set
by the {py:class}`Oracle <alf_core.oracle.oracle.Oracle>`'s scorer. This means a method validated offline moves to a real experiment without
rewriting it.

| Mode | Oracle scorer | Use case |
|------|---------------|----------|
| **{term}`Offline`** | a {py:class}`Dataset <alf_core.dataset.base_dataset.BaseDataset>` (held-out, pre-scored pool) | benchmarking and method development; reproducible, no external calls |
| **{term}`Online`** | a trained {py:class}`Model <alf_core.model.base_model.BaseModel>` (or simulator) | driving a real experiment where labels come on demand |

## Steps

Keep the `Dataset`, {py:class}`Surrogate <alf_core.surrogate.surrogate.Surrogate>`, {py:class}`Optimizer <alf_core.optimizer.optimizer.Optimizer>`, and {py:class}`Task <alf_core.tasks.base_task.BaseTask>` the same, and swap only the
`Oracle`'s scorer:

```python
# Offline: score against the dataset's held-out labels
oracle = Oracle(scorer=dataset)

# Online: score with a trained model on demand
oracle = Oracle(scorer=trained_model)
```

See [Intro to Active Learning](../explanation/intro-to-active-learning.md) for the concept.

## Reference notebooks

- [Offline Design Tutorial](https://github.com/instadeepai/alf/blob/main/tutorials/experiments/offline_design_tutorial.ipynb)
- [Online Design Tutorial](https://github.com/instadeepai/alf/blob/main/tutorials/experiments/online_design_tutorial.ipynb)