NeMo Data Designer

What NeMo Data Designer is

NeMo Data Designer is a synthetic data generation system that lets users set up dataset columns and models, preview sample records, and create a dataset. The framework handles batching, parallel execution, validation, statistical diversity, field correlations, and repeatable workflows. Listed data types include tabular data, text, code, structured JSON, images, and embeddings; output options include CSV, JSON, JSONL, and Parquet.

Users can choose an open-source Python library or a NeMo Microservice with REST API methods for previewing and creating datasets. Documented LLM endpoints include NVIDIA API Catalog, OpenAI, Azure OpenAI, Anthropic, vLLM, TGI, TensorRT-LLM, and OpenAI-compatible servers. The microservice also integrates with NeMo Inference Microservices, NeMo Customizer, and NeMo Evaluator. Data Designer does not host models, so users must provide LLM endpoints and inference capacity. The hosted trial requires an API key; jobs default to 10 records and allow up to 100 records per job. The default SECURE Jinja renderer includes sandboxing, context sanitization, template restrictions, and output guards.

Who it suits

The open-source library is positioned for developers with LLM access who want flexibility and customization. The microservice is aimed at teams already using NeMo Microservices.

Limits

The hosted free trial requires an API key and allows up to 100 records per job; jobs default to 10 records. Data Designer does not host models, so users must provide LLM endpoints and provision inference capacity.

Getting started

1. Visit https://build.nvidia.com/nemo/data-designer.
2. Choose the open-source Python library or the NeMo Microservice route.
3. For the hosted trial, obtain an API key.
4. Provide LLM endpoints and provision inference capacity.
5. Configure dataset columns and models, then preview samples.
6. Create a dataset and select an available output format.

Verdict

NeMo Data Designer offers both a customizable Python library and a NeMo Microservice route for synthetic data generation. Users need to arrange their own model endpoints and inference capacity.