NeMo Data Designer is an open-source framework for describing and generating multimodal synthetic datasets through a configuration file rather than a collection of one-off scripts. Each column can produce text, code, structured values, images, embeddings or statistically sampled data.
Dependencies between columns are resolved automatically, and the runtime schedules calls to model endpoints supplied by the user. A plugin system allows additional field types and processing logic. Because the full specification is stored as an inspectable artifact, teams can share the recipe, review how each value is created and rerun it after changing a model or requirement.
The workflow includes a preview-and-revision stage. Users first generate a small number of records, inspect quality and diversity, then revise the configuration before starting an expensive full run. The framework also retries failed requests. The authors describe uses spanning structured, agentic and image data, including datasets used for Nemotron development and enterprise deployments.
A reproducible generator does not guarantee that its outputs are accurate, unbiased or suitable for training. Quality checks still need to cover distributions, duplicates, unsupported claims and leakage from source models. The frameworkâs practical contribution is to make those choices visible and repeatable, giving reviewers something more concrete to audit than prompts scattered across notebooks and application code.