hub/pr_2521/en/datasets-data-designer.md

Data Designer

Data Designer is NVIDIA NeMo's framework for generating high-quality synthetic datasets using LLMs. It enables you to create diverse data using statistical samplers, LLMs, or existing seed datasets.

Prerequisites

pip install data-designer

Download datasets from the Hub as seeds

Use HuggingFaceSeedSource to load datasets directly from the Hub as seed data for generation.

import data_designer.config as dd
from data_designer.interface import DataDesigner

data_designer = DataDesigner()
config_builder = dd.DataDesignerConfigBuilder()

# Load seed data from HuggingFace
seed_source = dd.HuggingFaceSeedSource(
    path="datasets/gretelai/symptom_to_diagnosis/data/train.parquet",
    token="hf_...",  # Optional, for private datasets
)
config_builder.with_seed_dataset(seed_source)

# Reference seed columns in prompts
config_builder.add_column(
    dd.LLMTextColumnConfig(
        name="physician_notes",
        model_alias="openai-gpt-5",
        prompt="Write notes for a patient with {{ diagnosis }}. Symptoms: {{ patient_summary }}",
    )
)

preview = data_designer.preview(config_builder, num_records=5)

Push generated datasets to the Hub

Use the built-in push_to_hub method to upload generated datasets to the Hub.

# Generate dataset
results = data_designer.create(config_builder, num_records=1000, dataset_name="my-dataset")

# Push to Hub
url = results.push_to_hub(
    repo_id="username/my-synthetic-dataset",
    description="Synthetic dataset generated with Data Designer.",
    tags=["medical", "notes"],
    private=False,
)

Resources

• Data Designer Documentation
• GitHub Repository
• Seed Datasets Guide
• Guide to using Data Designer with Inference Providers