April 2025

2025.4.29

• Fix: Tweak the calculation of Deep Structure Stability in the SQS report for more consistent scores.

• Feature: Added new task for running the full sample-to-dataset pipeline.

• Feature: Adds support for users to provide their own test holdout dataset for safe synthetic data evaluation.

• Feature: Adds support for subcategory type sampling columns to the magic interface.

• Feature: Train and holdout dataset splits are now accessible as CSV outputs.

• Fix: Add display_name as an additional field to v2 report headers.

• Feature: Addition of data_designer_object.info.model_suite to display info of model suite selected for the current data designer object

• Fix: Model configs provided are now also validated upon calling data_designer_ojbect.validate()

• Fix: In cases when a person object has phone_number = None, None will be a NoneType rather than a string, consistent with how we treat email_address.

• Feature: A new "reasoning" model alias is now available for use within each model suite in data designer.

• Fix: Fixes a bug with the TextFT task config not picking up certain generation parameters

• Feature: Add configuration option to specify entity list for use in Transform NER functions.

• Fix: Format change to evaluation report.

• Task: Enforce a minimum of 200 input records or 10 holdout records for performing the holdout step in Workflows.

• Fix: Fixes bug where int holdout values greater than 1 would be cast to floats, causing holdout sampling to fail.

• Fix: Fixes two small typing related bugs in report column view and AIA calculations.

• Fix: Fixes a bad call on a default empty dictionary in the dataset designer report calculations.

2025.4.21

• Task: Fully wires in support for Transform sections of Multimodal report. Also updates the safe synthetics evaluate task to detect if transform and/or synthesis tasks were run in the workflow.

• Feature: Add optional Transform rule "description" field.

• Fix: Fixes an issue with the format for writing a report file in the client

• Feature: Adds a status field to the workflow run object

• Fix: Fix transform progress log off-by-one error

2025.4.16

• Feature: Free-tier users will now be given a one time allocation of 50 credits.

• Feature: Can now properly edit sampling configurations when using magic.add_sampling_column

• Feature: For the US locale, you can now filter on state.

• Bugfix: Support pandas Int64 and numpy bool_ dtypes for input columns for TabularFT

• Feature: New validate method on the DataDesigner object, which performs a task-level validation (by attempting to build the workflow) and a wholistic semantic validation of all the column definitions.

• Task: This changes the error messages users see, but not fundamentally a new feature or something that operates differently.

• Fix: Reduce the of assistant-like responses for the apache-2.0 suite.

• Fix: Expression columns now appear in the table when you call display_sample_record.

• Feature: Introduces designer.magic.refine_prompt, a tool for quickly iterating on LLM generation column prompt templates. Vary an existing prompt while keeping its intent with designer.magic.refine_prompt("my_column"). Instruct a specific change to the prompt with deisgner.magic.refine_prompt("my_column", "Ensure that the outputs are in Old English.").

• Task: Updates the multimodal report to show more relevant v2 workflow identifier info in the overview section.

• Fix: The error message for workflow step failure has been better standardized.

• Feature: Adds designer.magic.extend_columns(name, n) which can be used to increase the number of possible sampling values for a category.

• Fix: Fixes an issue with hybrid workflows not properly authenticating

• Fix: Remove experimental warnings from magic.refine_prompt and magic.extend_category.

• Fix: Minor update to how generation summaries are logged to the user with verbose_logging=True in AIDD.

• Feature: Release note: Update AIDD apache-2.0 suite to use newer models

• Fix: Magic llm columns now show up in aidd repr and display_sample_record.

• Fix: Fix for crashing bug preventing structured data references in Jinja templates for generated structured columns in Data Designer.

• Task: Renames the evaluate_dd_dataset task to evaluate_data_designer_dataset. Addresses serialization issues/requirements.

• Fix: Word-wraps REPR HTML displays to make notebook display of Data Designer columns more legible.

2025.4.9

• Fix: Change behavior of magic.add_sampling_column and magic.add_column to update existing columns (where possible) rather than raising an error.

• Fix: Improve error and status reporting for DataDesigner magic operations.

• Fix: Fix Text Metrics Failure for Multi-Metrics Report.

• Fix: Improve error reporting for Jinja specification errors when using DataDesigner.

• Fix: Any code language can now be specified for LLM generation columns in DataDesigner.

• Feature: Rename task from evaluate_ss_dataset to evaluate_safe_synthetics_dataset.

• Fix: Address an issue with jobs failing due to heartbeat issues.

• Feature: Add support for the following Jinja filters when writing expressions or prompt templates: abs, capitalize, first, float, int, items, last, length, list, lower, max, min, random, replace, reverse, round, sort, string, title, trim, truncate, unique, urlencode.

• Fix: Fix a bug in which workflows fail validation when globals are not provided.

• Fix: Fix for null values when attempting to reference LLM-generated structured data fields with ExpressionColumn.

• Feature: An error will now be thrown if the user tries to add a column with the same name as a latent person sampler.

• Feature: Person samplers now have an optional age_range argument, which defaults to adult ages.

• Fix: Fix a bug that crashed Data Designer jobs when sequence of Structured Output -> Expression -> Expression/LLMPrompt existed in the workflow.

2025.4.3-3

• Fix: Internal config updates.

2025.4.3-2

• Fix: Internal config updates.

2025.4.3-1

• Fix: Internal config updates.

2025.4.3

• Feature: Two new samplers have been added to AI Data Designer: Bernoulli Exponential and Bernoulli X (where X is any Scipy distribution).

• Feature: ModelAlias.NATURAL_LANGUAGE has been renamed to ModelAlias.TEXT.

• Feature: The AI Data Designer object __repr__ method now includes all column types.

• Feature: Enable stop-params in Default TabFT Config for improved inference efficiency.

• Fix: AIDD requirements update in gretel-client.

• Fix: Fix a regression with use_structured_generation.

• Fix: Support downloading CSV version of dataset task outputs.

• Fix: Fix a warning when importing the gretel_client in Python 3.12.

• Fix: Correct unicode handling for parsed outputs in Magic and Structured column generation.

• Fix: Bernoulli and BernoulliMixture will now work client side. This required a bug fix in a helper function.

• Fix: Update the print displays of column configurations when using magic to be copy-pasteable to facilitate easier editing.

• Fix: Fix a bug where pre-existing seed dataset columns would be dropped from the DataDesigner object after DataDesigner.magic.add_sampling_columns was used.

• Fix: Remove extraneous enum warnings from DataDesigner preview operations.

• Fix: Fix a bug where the DataDesigner's model_suite wasn't being respected when using magic to add LLM generation columns.