NeMo Data Designer
NeMo Data Designer is an open-source multimodal synthetic data generation framework released by NVIDIA. It adopts a declarative configuration format and supports generation of multiple data types including text, code, images, and embedding vectors, with built-in dependency resolution, scheduling, and fault tolerance mechanisms, targeting large model training data construction scenarios.
Timeline (last 90 days)
NVIDIA NeMo Data Designer (NDD) is an open-source, general-purpose multimodal synthetic data generation framework.
NDD adopts a declarative configuration format, in which users describe the desired data by defining each column of the dataset.
The column types supported by NDD cover text, code, structured output, images, embedding vectors, and statistical samplers.
NDD builds a preview-revise loop into its core workflow, allowing users to first generate a small number of records for inspection before re-running at full scale.
At runtime, NDD automatically resolves inter-column dependencies, schedules calls to user-defined model endpoints, and retries failed requests.
NDD's inter-column dependency resolution essentially constructs a directed acyclic graph (DAG) from which it infers which columns can be generated in parallel and which must wait for upstream columns to complete.
NDD provides a flexible plugin system that allows additional column types and functionalities to be introduced via plugins.
The NDD-related paper presents cross-scenario case studies covering structured tasks, agentic tasks, multimodal tasks, and domain-specific tasks.
Datasets used in the development of Nemotron models as well as datasets actually used in production-level enterprise deployments are included in NDD's case studies.
All Facts (9)
NVIDIA NeMo Data Designer (NDD) is an open-source, general-purpose multimodal synthetic data generation framework.
50%UnverifiedNDD adopts a declarative configuration format, in which users describe the desired data by defining each column of the dataset.
50%UnverifiedThe column types supported by NDD cover text, code, structured output, images, embedding vectors, and statistical samplers.
50%UnverifiedNDD builds a preview-revise loop into its core workflow, allowing users to first generate a small number of records for inspection before re-running at full scale.
50%UnverifiedAt runtime, NDD automatically resolves inter-column dependencies, schedules calls to user-defined model endpoints, and retries failed requests.
50%UnverifiedNDD's inter-column dependency resolution essentially constructs a directed acyclic graph (DAG) from which it infers which columns can be generated in parallel and which must wait for upstream columns to complete.
50%UnverifiedNDD provides a flexible plugin system that allows additional column types and functionalities to be introduced via plugins.
50%UnverifiedThe NDD-related paper presents cross-scenario case studies covering structured tasks, agentic tasks, multimodal tasks, and domain-specific tasks.
50%UnverifiedDatasets used in the development of Nemotron models as well as datasets actually used in production-level enterprise deployments are included in NDD's case studies.
50%