NeMo Data Designer Synthetic Data Skill
Build synthetic datasets and declarative data-generation pipelines from a natural-language description.
The skill explicitly requires permission before installing dependencies and requesting permission to retry network operations; its skill card also warns against exposing secrets in prompts, logs, or output. Deductions apply because it can generate names, addresses, and other potentially personal data and may call external inference services without fully documenting data flows, credential scope, isolation, or rollback.
The documentation defines interactive and autopilot branches, missing-CLI handling, network-failure feedback, and local-versus-cluster distinctions. Deductions apply because the key workflow files are not included, dependency and API compatibility are not pinned, and no tests cover the skill's key paths; static review cannot support a higher score.
Trigger conditions, target users, and the Python-script output format are reasonably clear, and the skill describes boundaries around seed data, person sampling, and cluster submission. Deductions apply because non-fit scenarios, Chinese-language support, and mainland-China reachability are not addressed, while core functionality may depend on external model services.
SKILL.md includes metadata, dependency guidance, references, troubleshooting, and an output template; the skill card provides Apache 2.0 licensing, ownership, version 0.2.0, and risk guidance. Deductions apply because author identity, changelog, explicit maintenance responsibility/update path, and examples are incomplete; the benchmark also reports missing recommended Instructions and Examples sections.
The stated goal is to produce a directly usable Data Designer configuration script, with concrete requirements for the loader function, inline metadata, and model configuration; the evaluation case specifies expected fields and file output. Deductions apply because no generated result or third-party execution evidence is supplied, so quality, compatibility, and benefit over manual alternatives cannot be verified statically; the static cap is 7.
The files provide concrete commands, configuration patterns, an evaluation case, and a static-validation report, giving limited auditability. Deductions apply because evaluation tasks and Tier 3 details are unavailable, all reported metrics are N/A, and there is no skill-specific test suite or independent reproduction evidence; the static cap is 5.
- Before execution, confirm data-designer, Python version, model-provider, network, and credential configuration; do not place real personal data or secrets in seed data, prompts, logs, or generated outputs by default.
- The key workflow files and actual execution results are not provided; validate locally in an isolated environment and manually review generated data for privacy, bias, quality, and licensing suitability.
- Mainland-China reachability is unverified; external inference services or platform resources may require additional network configuration.
What does this skill do, and when should you use it?
This skill is for users who want to create a dataset, generate synthetic data, or build a data-generation pipeline. It selects Interactive or Autopilot mode based on whether the user wants to answer configuration questions. The workflow writes a Python file in the current directory with a load_config_builder() function that returns a DataDesignerConfigBuilder. It fits teams that want structured, declarative synthetic-data generation while accounting for environment setup, model configuration, and network availability.
Accepts a dataset description; selects Interactive or Autopilot mode; reads only the workflow matching that mode, plus conditional references for person sampling, seed datasets, or NeMo Platform plugin details; writes a descriptive Python configuration file with PEP 723 dependency metadata; builds a synthetic-data configuration with NeMo Data Designer's DataDesignerConfigBuilder; supports declaring model configurations in scripts when LLM columns are used; keeps output columns by default; and applies documented rules for sampler parameters, Jinja2 templates, and nested LLM-judge scores.
- A data engineer wants to generate synthetic customer reviews, conversations, or other structured records from a natural-language specification.
- A machine-learning team needs to prototype a dataset-generation pipeline for training or evaluation.
- A user who prefers minimal interaction wants the agent to make reasonable assumptions and produce a generation script.
- A team building from existing records explicitly provides seed data and needs the corresponding seed-dataset workflow.
What are this skill's strengths and limitations?
- Focused on synthetic dataset creation and data-generation pipelines.
- Provides both Interactive and Autopilot workflows.
- Documents important rules for column retention, sampler parameters, Jinja2 templates, and LLM-judge score access.
- Script-defined model configurations are portable between local run and cluster submit.
- Requires Python >= 3.11 and a NeMo Data Designer environment.
- LLM-based columns generally require an available model configuration or inference provider; the SKILL.md does not specify a concrete model.
- The supplied source does not include the full workflow files, installation procedure, or ready-made dataset templates.
- Preview may fail in a network-restricted sandbox, and the skill leaves retry or installation decisions to the user.
How do you install this skill?
Install the specific skill with the skills CLI command documented by the repository README:
npx skills add nvidia/skills --skill nemo-data-designer-plugin --yes
The README does not specify the exact installation directory; the skill becomes available the next time the agent loads skills and encounters a relevant task.
How do you use this skill?
Prompt the agent with a dataset request such as: “Create a synthetic customer-review dataset with customer name, city, product, rating, and review.” Interactive mode is the default. Use wording such as “be opinionated,” “you decide,” or “just build it” to select Autopilot mode. The result should be a descriptive Python file in the current directory containing load_config_builder().