NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training
by NVIDIA
Generate validated synthetic training and evaluation data from samplers, LLMs, seed datasets, and agent traces.
Overview
NeMo Data Designer is NVIDIA's open-source (Apache-2.0) framework for generating synthetic datasets, from scratch or from seed data. You declare columns (samplers, LLM-generated text, structured outputs, images, validators), preview samples, then generate at scale while the library handles batching, parallelism, and validation against Python, SQL, custom validators, or LLM judges. The concepts section covers columns, seed datasets, model providers, MCP tool use, processors, workflow chaining, agent rollout ingestion, and performance. Seven tutorial notebooks run in Google Colab and go from a first dataset through Jinja prompts and structured outputs, external seed data, image context, image generation, and image editing. The recipes are the most useful part for practitioners: text-to-Python and text-to-SQL pairs, multi-turn chat, PDF question answering over MCP, agent trace distillation for Claude Code, Codex, and Hermes formats, and a nine-recipe pipeline for visual QA on long PDFs. The 19 dev notes describe how NVIDIA built datasets for its Nemotron models, including an enterprise text-to-SQL set, and how the library handles async execution and adaptive concurrency. It runs as a Python library or a NeMo Microservice. Version 0.10.0 shipped on November 2, 2026, and dropped Python 3.10. The repository had 2.4k stars and 246 forks on November 11, 2026. A December 2025 Hacker News post announcing the open-source release drew 8 points and one comment, and independent tutorials were hard to find.
At a Glance
- Topic
- Fine-Tuning
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- Self-paced: seven tutorial notebooks runnable in Google Colab; NVIDIA publishes no time estimate
- Provider
- NVIDIA
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Declare sampler, LLM, structured-output, and validator columns to generate a first synthetic dataset
- ✓Build conditional prompts with Jinja expressions and schema-constrained outputs for generated records
- ✓Seed synthetic data generation from an existing external dataset instead of starting empty
- ✓Configure model providers, inference parameters, and concurrency for generation at larger scale
- ✓Give generation columns MCP tool access to produce grounded, tool-using training examples
- ✓Score generated records with Python, SQL, custom validators, and LLM judge columns
- ✓Distill Claude Code, Codex, and Hermes agent traces into structured workflow records
- ✓Chain workflows to pause at a review stage, export data, and resume
Highlights
- •Recipes and dev notes show how NVIDIA built Nemotron datasets, including a text-to-SQL set scored by five judges on 15 dimensions
- •A nine-recipe pipeline generates visual QA training data from long PDFs, from OCR through a stronger-model judge filter
- •Agent rollout ingestion converts ATIF, Claude Code, Codex, and Hermes traces into structured records
- •A Retriever SDG plugin turns documents into retriever training and BEIR evaluation data
- •Tutorials run in Google Colab, so you can try the library without local setup
Who It’s For
Best For
- ✓ML engineers building fine-tuning datasets when real data is scarce or restricted
- ✓Teams creating evaluation sets for RAG, text-to-SQL, or tool-calling agents
- ✓Post-training researchers who need reproducible synthetic data pipelines
Prerequisites
- •Python 3.11 to 3.14 for a local install, or a Google Colab account for the tutorials
- •An LLM endpoint or API key, such as NVIDIA Build API, OpenAI, or OpenRouter
- •Enough fine-tuning or evaluation experience to specify the data your model needs
FAQ
What is NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training?
NeMo Data Designer's documentation teaches ML engineers how to generate synthetic datasets for fine-tuning and evaluation, starting from nothing or from seed data. After the tutorials and recipes, you can design correlated columns, validate records with code or LLM judges, generate tool-using agent examples, and run the workflow at larger scale.
Is NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training free?
NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training is free to access.
What level is NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training for?
NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training is aimed at a intermediate audience. Recommended background: Python 3.11 to 3.14 for a local install, or a Google Colab account for the tutorials, An LLM endpoint or API key, such as NVIDIA Build API, OpenAI, or OpenRouter, Enough fine-tuning or evaluation experience to specify the data your model needs.
How long does NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training take?
Expect roughly Self-paced: seven tutorial notebooks runnable in Google Colab; NVIDIA publishes no time estimate. Most learners work through it at their own pace.
What will I learn from NeMo Data Designer Documentation: Synthetic Data Generation for LLM Training?
You'll learn: Declare sampler, LLM, structured-output, and validator columns to generate a first synthetic dataset; Build conditional prompts with Jinja expressions and schema-constrained outputs for generated records; Seed synthetic data generation from an existing external dataset instead of starting empty; Configure model providers, inference parameters, and concurrency for generation at larger scale; Give generation columns MCP tool access to produce grounded, tool-using training examples; Score generated records with Python, SQL, custom validators, and LLM judge columns; Distill Claude Code, Codex, and Hermes agent traces into structured workflow records; Chain workflows to pause at a review stage, export data, and resume.
Topics
Sources
This page was written from 7 sources, 3 on domains other than docs.nvidia.com.