NeMo Data Designer

NeMo Data Designer is an open-source Python framework from NVIDIA for generating synthetic datasets from scratch or from seed data. You connect LLM endpoints such as NVIDIA Build, OpenAI, OpenRouter or vLLM, and the library handles batching, parallelism and validation. The latest release listed on GitHub is v0.9.4.

What It Is

Data Designer is a synthetic data generation library. You define a dataset as a set of columns, using samplers, LLM text columns, image models or seed datasets. Columns can depend on one another, so fields in a record can be correlated. The documentation contrasts this with raw LLM calls, saying it provides statistical diversity, field correlations, automated validation and reproducible workflows.

Workflow

The documented loop is Configure, Preview, Create. You configure columns and models through a Python config builder, preview sample records to iterate, then create the full dataset at scale. Models and providers can be set up with the data-designer config CLI commands. Local Python config modules can also accept workflow-specific command-line arguments.

Capabilities

• Generate text, structured data and images, with image, audio and video context for multimodal workflows.
• Connect generation to local or remote MCP servers and capture interaction traces.
• Validate and score outputs with Python, SQL, custom validators and LLM judges.
• Sample realistic person data with demographic attributes.
• Chain workflows, and pause, inspect and resume runs.
• Extend it with plugins for custom columns, seed readers and processors.

The documentation includes recipes for text-to-SQL, text-to-Python, multi-turn chat, image generation, document QA and agent trace distillation.

Agent Skill and Telemetry

A data-designer skill for coding agents lets you describe the dataset you want while the agent handles schema design, validation and generation. It is tested with Claude Code and Codex and installs via npx skills add NVIDIA-NeMo/DataDesigner. The library includes optional anonymous telemetry covering model names and token counts, which can be disabled with NEMO_TELEMETRY_ENABLED=false. The README notes that NVIDIA Build is intended for evaluation and testing only.