June 2025

2025.6.17

• Fix: Allows skipping artifact validation during gretel_client import

• Feature: - New with_synthetic_personas boolean argument as been added to person samplers.

• Task: The create method in AI DataDesigner and Safe Synthetics will now consistently group workflow runs under a previously created workflow based on the name provided, unless the new_workflow boolean flag is explicitly set to True.

For example:

# providing no name creates a new workflow (ex, with id=1)
aidd_or_ssd_obj.create()

# re-running the above re-uses the workflow created above. New run is added to workflow with id=1
aidd_or_ssd_obj.create()

# giving a new name results in a new workflow (ex, with id=2)
aidd_or_ssd_obj.create(name="apples")

# re-running with a used name results in using the previously created workflow (ex, with id=2). A new run is added to workflow with id=2
aidd_or_ssd_obj.create(name="apples")

# changing the name again to something else will result in a new workflow (ex, with id=3)
aidd_or_ssd_obj.create(name="bananas")

# switching back to not providing any names will result in using the the last workflow used (ex, with id=3 created for bananas)
aidd_or_ssd_obj.create()

# at any time I can force a new workflow creation regardless of any state maintained (ex, with id=4)
aidd_or_ssd_obj.create(new_workflow=True)

# forcing a new workflow using a name previously used will reset the workflow id saved for that name in the state (ex, with id=5, for apples)
aidd_or_ssd_obj.create(new_workflow=True, name="apples")

# any new workflows created with the namne "apples" without the new_workflow flag will now be under the new workflow with id=5
aidd_or_ssd_obj.create(name="apples")

This behavior is tied to a specific Gretel object. As long as the same Gretel object is used this behavior is preserved regardless of whether you create new instances of data designer of safe synthetics objects within the same execution environment.

2025.6.10

• Feature: - Allow the use of the {{ ... | escape }} and {{ ... | forceescape }} filters in Jinja templates.

• Fix: Makes the pycryptodome dependency optional when importing gretel_client

• Fix: - bring-your-own model suite now allows multiple model configs with different aliases pointing to the same model but with different generation parameters. This allows one to control for different model behavior in the same workflow.

2025.6.3

• Fix: - Enable training without unsloth when rope_scaling_factor > 1.