Tutorials

These tutorials demonstrate how to build Data Designer configurations and execute them through the NeMo Data Designer plugin.

The code snippets on this page are for conceptual demonstration purposes only. For runnable examples, jump ahead to the Basics or Seeding tutorial.

Configuration and Execution

Data Designer separates configuration (building dataset schemas) from execution (generating the data).

Part 1: Build Configs (Library)

Use data_designer.config to define your dataset. See the library documentation for comprehensive guides on column types, constraints, and processors.

import data_designer.config as dd
config_builder = dd.DataDesignerConfigBuilder(model_configs)
config_builder.add_column(dd.SamplerColumnConfig(...))
config_builder.add_column(dd.LLMTextColumnConfig(...))

Part 2: Execute (Plugin)

Run the configuration with the CLI or SDK:

nemo data-designer preview product_reviews.py --num-records 5
nemo data-designer create product_reviews.py --num-records 30

import os
from nemo_platform import NeMoPlatform
client = NeMoPlatform(
    base_url=os.environ.get("NMP_BASE_URL", "http://localhost:8080"),
    workspace="default",
)
data_designer = client.data_designer
preview = data_designer.preview(config_builder)
job = data_designer.create(config_builder, num_records=1000)

Prerequisites

Complete Setup to ensure you have the NeMo Services running locally and an inference provider available. These tutorials reference the default NVIDIA Build model provider, which is created as default/nvidia-build during setup.

Tutorials

The Basics

Generate a product review dataset using samplers and LLM-generated text. Learn the fundamentals of building configurations and executing jobs.

beginner data-designer

Seeding

Use external datasets to ground synthetic data generation. Generate realistic patient medical notes from symptom-to-diagnosis data.

intermediate data-designer

Generate Retrieval Training Data

Turn your own document corpus into judged question-and-answer pairs, ready for embedding fine-tuning. Hard-negative mining uses the current encoder to select non-answer passages that rank near each positive, giving the model useful examples to distinguish.

intermediate data-designer retrieval