nvidia-nemo/datadesigner — explained in plain English
Analysis updated 2026-05-18
Generate a large, varied synthetic dataset to train or test a machine learning model without real user data.
Turn a small set of real sample records into a much larger dataset with similar patterns.
Build a dataset with fields that depend on each other, like a product category and a matching review.
Validate and score generated data automatically before using it downstream.
| nvidia-nemo/datadesigner | facebookresearch/ai4animationpy | tencent-hunyuan/hy-world-2.0 | |
|---|---|---|---|
| Stars | 1,859 | 1,846 | 1,911 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | moderate | hard |
| Complexity | 3/5 | 4/5 | 5/5 |
| Audience | developer | researcher | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires an API key from a supported model provider such as NVIDIA Build, OpenAI, or OpenRouter.
NeMo Data Designer, from NVIDIA, is a Python tool for creating synthetic datasets, meaning data that is generated on purpose rather than collected from the real world. You can build a dataset entirely from scratch, or start from your own sample data and grow it into something larger and more varied. Instead of just asking a language model to write rows of data in one shot, Data Designer lets you build a dataset column by column, with control over how each column behaves. Some columns can pull values from statistical samplers, such as picking randomly from a fixed list of categories. Other columns can be generated by a language model using a prompt you write, and that prompt can reference the values already generated in other columns, so fields stay meaningfully connected to each other rather than being random and disconnected. The tool also includes validators, which check that generated data actually meets the rules you care about, using plain Python code, SQL queries, or custom checks you write yourself. On top of that, you can use a language model as a judge to score how good each generated record is, which helps catch low quality output before it ends up in your final dataset. A preview mode lets you generate a small sample first, so you can check the results look right before running a full, larger generation job. Getting started involves installing the package, setting an API key for a supported model provider such as NVIDIA Build, OpenAI, or OpenRouter, and then writing a short Python script that defines your columns and generates a preview. There is also a command line interface for configuring providers and models without writing code. The project ships as a package for coding agents as well, so an agent like Claude Code can be asked to design a dataset in plain language and have the tool handle schema design, validation, and generation on its own. Data Designer is released under the Apache 2.0 license and collects optional anonymous telemetry about which models are used, which can be turned off.
NeMo Data Designer is an NVIDIA Python tool for generating synthetic datasets from scratch or seed data, with controllable, validated, and quality scored columns.
Mainly Python. The stack also includes Python, LLMs, CLI.
You can use, modify, and distribute this software freely, including for commercial purposes, as long as you follow the Apache 2.0 license's attribution and notice requirements.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.