NVIDIA's NeMo Data Designer Generates Synthetic Training Datasets at Scale
NVIDIA open-sources NeMo Data Designer, a Python framework that generates production-grade synthetic training datasets with built-in validation, field correlations, and LLM-as-a-judge scoring.
Takeaways
- NVIDIA open-sources NeMo Data Designer, a Python framework for generating production-grade synthetic training datasets from scratch or seed data.
- Uses a declarative column-based schema -- define fields, types, relationships, and validation rules; the engine handles LLM orchestration and dependency resolution.
- Supports NVIDIA Build, OpenAI, OpenRouter, vLLM, and Hugging Face endpoints via OpenAI-compatible APIs; exports to CSV, JSON, and Parquet.
- Includes LLM-as-a-judge scoring, built-in validators (Python, SQL, remote), and a preview mode to iterate before full-scale generation.
- Ships a coding agent skill for Claude Code -- describe the dataset you want in plain English and the agent writes the schema config for you.
- Free under Apache 2.0; install with
pip install data-designer. GitHub repo | Docs
Getting quality training data is one of the most stubborn bottlenecks in applied ML. Three pressures have converged to make synthetic data generation a mainstream concern: frontier models have made high-quality generation cheap, real data scarcity has become a genuine constraint, and the most useful domain-specific corpora are often proprietary, legally encumbered, or simply too small to train on. NVIDIA's answer to this is NeMo Data Designer, an open-source Python framework for generating synthetic training datasets at scale.
Beyond "just prompt an LLM"
Data Designer goes beyond simple LLM prompting. Whether you need diverse statistical distributions, meaningful correlations between fields, or validated high-quality outputs, it provides a flexible framework for building production-grade synthetic data. The key design insight is declarative: users declare what their data should look like -- columns, types, relationships, validation rules -- and the engine figures out how to generate it.
Data Designer is a framework for orchestrating complex synthetic data generation workflows. It coordinates LLM calls, manages dependencies between data fields, handles batching and parallelization, and validates generated data against specifications. Think of it as a pipeline orchestrator where your schema is the config and LLMs are the workers.
What it actually does
The framework is built around a column-based data model. You define each column as a typed config -- statistical samplers for categorical or numeric fields, LLM-powered text columns with Jinja-style templating, or columns derived from other columns. The engine resolves the dependency graph and generates everything in the right order.