NVIDIA NeMo Data Designer iconNVIDIA NeMo Data Designer

Open-source NVIDIA framework for generating synthetic datasets with LLMs, samplers and validators; carries on the synthetic data tech from Gretel

2.4k+ GitHub Stars
Apache-2.0 License
Gretel Tech Lineage

Overview

NeMo Data Designer is NVIDIA's open-source (Apache-2.0) Python framework to "generate high-quality synthetic data from scratch or from seed data". You declare the columns you want, and it fills them from statistical samplers, LLMs, image models or a seed dataset, keeps dependencies between fields, and checks each row with Python, SQL or LLM-judge validators. It supports text, structured data and images, can call MCP servers, and ships an agent skill for coding agents such as Claude Code and Codex. Models come from NVIDIA Build, OpenAI, OpenRouter or your own endpoints (NVIDIA Build is for evaluation only). NVIDIA reportedly acquired the synthetic data startup Gretel in March 2025; gretel.ai now redirects to NVIDIA, its open-source library is archived, and a Gretel co-founder says Data Designer includes "the synthetic data tech my team originally built at Gretel". Also from NVIDIA: NVIDIA Nemotron, whose training data Data Designer helps build.

The Verdict

Who Should Use NeMo Data Designer?

Best For

  • ML teams generating fine-tuning, evaluation or test data with LLMs
  • Former Gretel users looking for the successor
  • Engineers who want synthetic data as code they can version

Not Ideal For

  • Business users who want a no-code synthetic data app
  • Teams that need a vendor-supported hosted product today

What's Great

  • Free, permissive license and very active repo
  • Validators and LLM judges built into the pipeline
  • Works with many model providers, not only NVIDIA

Watch Out For

  • A library, not a hosted app: you run and pay for the models
  • NVIDIA Build endpoints are for testing, not production
  • Gretel's old console and docs are gone

Pricing

View all features & details

Key Features

  • Samplers, LLM and image-model columns
  • Seed datasets and field dependencies
  • Python, SQL and LLM-judge validators
  • MCP tool calls with traces
  • Preview, resume and run monitoring

Use Cases

  • Fine-tuning and instruction data
  • Evaluation and test sets for agents
  • Privacy-safe stand-ins for real records

How It Compares

Feature NeMo Data Designer Argilla
Main job LLM-driven synthetic data Human feedback and labeling
Open source Yes (Apache-2.0) Yes
Hosted product No Yes

Sources: NVIDIA synthetic data · Hacker News (Gretel co-founder) · gretel-synthetics (archived)

Is this your tool? Add a badge