Skip to main content
Version: Main branch

Dataset

  • An immutable snapshot of a query saved to db.artifact_store
  • Used (currently) for validating model performance
  • Snapshot allows users to know that their validation comparisons are apples-to-apples

Usage pattern

(Learn how to build a model here)

from superduper import Listener

ds = Dataset(
    'my-valid-data',
    select=db['my_table'].select(),   # `.select()` selects whole table
)

db.apply(ds)

Explanation

  • On creation superduper queries the data from the db.databackend based on the select parameter.
  • The data queries like this is saved as a persistent blob in the db.artifact_store.
  • When the dataset is reloaded, the select query is not executed again, instead the data is reloaded from the db.artifact_store. This ensures the Dataset is always "the same".
  • Dataset is handy for making sure model validations are comparable.