Nvidia introduced the NeMo Data Designer framework for synthetic data

Nvidia company released NeMo Data Designer (NDD) — an open framework for generating multimodal synthetic data. The tool is designed to create datasets that include text, code, structured answers, images, and embeddings.

NDD's architecture is based on a declarative configuration format, which allows users or AI agents to describe each dataset column in detail. The system supports flexible adjustment of statistical parameters to control data diversity. The built-in plugin system allows adding new data types and extending functionality for specific tasks.

The workflow in NDD is built on an iterative "review and revise" cycle. Users can generate small samples of records to check quality, adjust specifications, and run full-scale generation. At the execution stage, the framework automatically resolves dependencies, manages model API calls, and performs retry attempts in case of request failures.

Developers used NDD when creating models of the Nemotron family and in several corporate projects. The framework's configuration files are verifiable artifacts, which simplifies workflow sharing and ensures reproducibility of training results. The project's source code is distributed under the Creative Commons Attribution 4.0 license.

The implementation of such tools allows automating the preparation of high-quality training samples for specialized models. The use of synthetic data is becoming a standard in the industry, allowing to bypass limitations on access to labeled real data and accelerate the development of domain solutions.