Data augmentation

Augmentation is optional, but recommended when a dataset is small or lacks diversity. AutoScientist can add two independent types of rows before its training iterations begin:

• Domain rows add examples specific to the dataset’s domain. Existing rows count toward the recommended 20,000-domain-row target.
• General rows add diverse examples that help the model retain broader capabilities.

Augmentation here is part of the AutoScientist request. To improve and restructure the source dataset first, use Adaptive Data.

Calculate the additions

The following request targets at least 20,000 domain rows and adds 8,000 general rows:

from adaption import Adaption

client = Adaption()

dataset_id = "dataset_abc123"
dataset = client.datasets.get(dataset_id=dataset_id)

domain_rows_to_add = max(0, 20_000 - (dataset.row_count or 0))

run = client.autoscientist.create(
    dataset_id=dataset_id,
    augmentation_domain_rows=domain_rows_to_add,
    augmentation_general_rows=8_000,
)
print(run.id, run.status)

This produces up to 28,000 training rows: the existing dataset plus enough domain rows to reach 20,000, followed by 8,000 general rows. If the dataset already contains at least 20,000 rows, domain_rows_to_add is zero.

Omit either augmentation parameter to skip that type of addition. AutoScientist prices and sizes each requested expansion before training. See the create API reference for current limits and request fields.