SDV Enterprise 0.50 (Jul 14, 2026) - Extrapolating ranges

Release Notes
1

What’s New?

In this version, we’re releasing a new functionality for range extrapolation. By default, SDV synthesizers learn patterns from your training data, including ranges and category values. This feature allows you to expand those ranges when generating synthetic data.

The range extrapolation feature covers:
:star: Expanding the min and max boundaries for continuous attributes, like numerical or datetime
:star: Expanding to new category values that are not covered in the training data
:star: Adding support for nulls (or non-nulls) even if your training data doesn’t cover that scenario

The feature is especially useful when your training data represents only a small slice of your full database or if you want to simulate hypothetical scenarios that haven’t occurred yet.

Range extrapolation is available as part of the Targeted Sampling bundle.

Release History

0.50.0 (Jul 14, 2026): The initial release of v0.50 which includes the range extrapolation feature.

0.50.1 (Jul 21, 2026): This release encapsulates some small updates made to the underlying SDMetrics and RDT libraries that power SDV. In SDMetrics, the reports now accept a unified metadata dictionary (rather than a separate single- and multi-table metadata). In RDT, some categorical transformers can now be configured to ignore missing values. These are optional and do not result in any major user-facing changes.

0.50.2 (Jul 28, 2026): This release unifies the functions and parameters for the diagnostic and quality evaluations so you do not have to import separately for single vs. multi-table datasets.

0.50.3 (Aug 11, 2026): This release fixes a few bugs when generating Regexes for IDs, and addresses some warnings that were being produced by the differential privacy bundle.

0.50.4 (Aug 18, 2026): This release performs some routine maintenance. No user-facing changes were made.

0.50.5 (Aug 25, 2026): This releases fixes some bugs related to visualizations. It also introduces some utility functions that you can use to evaluate the privacy and anonymization of PII values.

3

Hi, I tried this for categorical type column and it worked. However I’m expected to list all the possible categories that are not present in the training set along with the ones present in the training set in the ranges json configuration. For example, my training set has “Red”, “Green”, “Blue”, “Yellow”, & “Black” under colors column and I had to input all those 5 along with new colors “White”, & “Purple” (total 7) in the json. Now what if there are 1000s of unique categorical values in the training set ?

4

Hi @Chakravarthy, this is a great Q. I will reply below with some thoughts. But I will also start a separate thread so that we can discuss in more detail about the particulars of your data.

In SDV, we consider a variable to be “categorical” when there are only a finite, preset number of possible values that the data is allowed to have. Many times, these are described in databases as ENUMS or possibly VARCHAR.

To extrapolate outside of the known categories (in the training data), it is necessary that SDV knows what all the possibilities are. I understand that inputting them all yourself may not be feasible if you’re looking to do this across many columns and projects. To solve this, team is also working on a few features:

  • If you’re using AI Connectors, the connector will be able to query for the full range/category information and create the JSON for for you
  • If you have CSVs, SDV Enterprise can create a JSON based on your training data. However, to expand upon the ranges from the training data, you would then have to go in and add values.

This will be available in the future and would solve the problem of manually inputting everything.

Is your data categorical?

You mention that there might be thousands of categorical values. In this case, I’d be curious if your data is best described as "categorical" or whether there is another sdtype label that might be make more sense for it. For example, perhaps it is a “real world concept” such as a credit_card_number, id, or something else? I will follow up via a separate thread.