NeMo Data Designer
Generate high-quality synthetic data from scratch or from your own seed data.
Evidence shows: README clearly states telemetry can be disabled (NEMO_TELEMETRY_ENABLED=false) and describes the data collected (model names and token counts), indicating data flow transparency. However, no clear evidence of least privilege, user confirmation, or rollback mechanisms. Dependency security: pyproject.toml comments mention security-fix versions (e.g., aiohttp, jupyter-server), showing some attention. External effects: README mentions configurable third-party endpoints but lacks detail. Source attribution: README provides citation info, but publisher identity is unverified. Deductions: lack of implementation details for least privilege, user confirmation, rollback, and sensitive data handling.
Evidence shows: Project structure is clear with multiple test files (e.g., fern/scripts/tests/), indicating good self-consistency. Dependency availability: pyproject.toml defines dependency groups but no full lockfile details (e.g., uv.lock). Failure messages: tests do not reveal explicit error handling or user-friendly failure messages. Deductions: insufficient evidence for failure messages and dependency availability.
Evidence shows: README describes multiple use cases (diverse data generation, relationship control, quality validation) and provides quick-start examples, indicating good audience and scenario coverage. Capability boundaries: docs mention async engine and model configuration but lack explicit limits. Trigger precision: CLI commands and skill installation provide triggers but not detailed conditions. Environment fit: supports Python 3.10-3.14 with installation instructions. Deductions: insufficient detail on trigger precision and capability boundaries.
Evidence shows: README provides clear information architecture including quick start, docs links, and contribution guide. Install notes are detailed (pip and source). Naming stability: no major naming changes observed but no version history. Examples and FAQ: code examples and docs links provided, but no FAQ. Known limitations: README mentions async engine caveats but not comprehensive. License: Apache-2.0, complete. Versioning/changelog: no CHANGELOG file found. Maintenance responsibility: contribution guide and issue tracker provided. Deductions: missing FAQ, known limitations, and changelog.
Evidence shows: Output usability: README describes generated datasets suitable for production-grade synthetic data, with preview mode. Marginal value: multiple generation methods (statistical sampling, LLM, seed data) but no comparison with alternatives. Cost-benefit: no performance or cost data provided. Deductions: insufficient evidence for cost-benefit and marginal value.
Evidence shows: Claims in README (e.g., '20T+ Tokens Processed') lack specific data sources but provide doc links. Cross-source corroboration: no independent verification found. Fact-inference separation: README distinguishes feature descriptions and telemetry data but does not explicitly separate facts and inferences. Deductions: claims lack traceability and no independent verification.
- Publisher identity is unverified; treat source with caution.
- Telemetry is enabled by default; be aware of privacy implications, disable with NEMO_TELEMETRY_ENABLED=false.
- Dependency security comments indicate known vulnerabilities in some dependencies, but no full lockfile is provided; consider using uv.lock for reproducibility.
- Performance claims in README (e.g., 20T+ tokens) lack verifiable data sources.
What does this agent do, and when should you use it?
NeMo Data Designer is NVIDIA's framework for generating synthetic datasets, either from scratch or based on existing seed data. It provides statistical samplers, LLM text columns, dependency-aware generation, built-in validators (Python, SQL, and custom remote), and LLM-as-a-judge scoring. It supports preview mode for rapid iteration before full-scale generation. Installation is via pip, and it supports multiple model providers (NVIDIA Build, OpenAI, OpenRouter). The framework offers a Python API and CLI tools, and includes a skill for coding agents, enabling integration with tools like Claude Code.
NeMo Data Designer generates synthetic data through column configurations. Users define columns using SamplerColumnConfig and LLMTextColumnConfig, specifying sampler types (e.g., CATEGORY) and prompt templates. It uses an async engine to process columns concurrently and supports dependency-aware generation to control correlations between fields. Built-in validators (Python, SQL, remote) ensure data quality, and LLM-as-a-judge scores outputs. Users can preview datasets via the DataDesigner class's preview method and configure model providers via the data-designer config CLI. The framework also offers a CLI tool and a skill integration that enables coding agents (e.g., Claude Code) to generate datasets from natural language descriptions.
- Data scientists needing diverse training data for ML models when real data is scarce or sensitive.
- Engineers generating synthetic datasets with specific field correlations, such as customer data for testing database queries.
- Researchers creating annotated datasets quickly to evaluate LLM outputs.
- Developers building multi-language text datasets using LLM prompt templates for benchmarking.
- QA teams validating data integrity and business rules before data enters production.
What are this agent's strengths and limitations?
- Seamless support for multiple model providers (NVIDIA Build, OpenAI, OpenRouter) with no code changes.
- Built-in validators (Python, SQL, remote) ensure generated data quality and business rule compliance.
- Async engine improves performance on large datasets by parallelizing columns and adapting concurrency.
- Preview mode and LLM-as-a-judge scoring enable rapid iteration before full-scale generation.
- Requires valid API keys (NVIDIA Build, OpenAI, or OpenRouter), which may incur costs.
- Data privacy concerns when using NVIDIA Build endpoint; its terms of service may restrict production use.
- The framework is still early-stage and may lack some documentation or stable APIs.
- Skill integration is primarily aimed at Claude Code; support for other coding agents is limited.
How do you install or deploy this agent?
Install via pip: pip install data-designer. Or from source: git clone https://github.com/NVIDIA-NeMo/DataDesigner.git && cd DataDesigner && make install. Requires Python 3.10 to 3.14.
How do you use this agent?
Set one API key: export NVIDIA_API_KEY="..." or export OPENAI_API_KEY="..." or export OPENROUTER_API_KEY="...". In Python, import data_designer.config and DataDesigner, create a DataDesignerConfigBuilder, add column configs (e.g., SamplerColumnConfig or LLMTextColumnConfig), then call data_designer.preview(config_builder=config_builder) and preview a sample.
FAQ
Does using NeMo Data Designer cost money?
Can I train on my own company data?
Will the generated synthetic data contain personally identifiable information?
How do I ensure data quality?
What if I encounter performance issues?
inference_parameters.timeout to match your actual per-request latency.