Data Designer Synthetic Data Skill
Build synthetic datasets and data-generation pipelines from a natural-language specification.
The skill limits workspace exploration, forbids installing dependencies without permission, and gives controls for seed data, personal-data sampling, sandbox retries, licensing, and security reporting. Points are deducted because external model/network data flows, synthetic-person privacy risks, rollback, and general confirmation before preview/create are not fully disclosed.
Interactive and autopilot workflows, missing-CLI handling, network-failure guidance, parameter inspection, and several common-error messages are specified. Points are deducted because there is no executable test suite for this skill, key behavior depends on an external CLI, model configuration, and context files not supplied here, and several failure paths can only be inferred statically; the static cap keeps this at or below 10.
The audience, dataset-description input, Python-script output, interactive/autopilot routing, and seed-data conditions are reasonably clear. Points are deducted for weak non-fit boundaries, model/hardware/cost/scale guidance, limited Chinese-language support, and uncertain mainland-China reachability where preview or generation may depend on overseas services.
The documentation includes goals, rules, workflows, tips, troubleshooting, an output template, and layered references; the skill card states owner, Apache 2.0 licensing, version v0.6.1, use case, and governance-related information. Points are deducted for no formal changelog, no explicit maintenance owner/update path, limited tabular examples and FAQ coverage, and the supplied metadata being marked NOASSERTION; the benchmark also reports missing recommended Instructions, metadata.author/tags, and run_script guidance.
The skill clearly targets a directly usable Python configuration file with load_config_builder() and covers samplers, person data, validation, preview, and creation. The benchmark reports positive correctness and effectiveness results for the core task. Points are deducted because the evaluation dataset and reproducible artifacts are unavailable, outputs still require review, and success depends on external CLI, model, and configuration availability; the static cap keeps this at or below 7.
The pinned source material provides workflows, references, an example script, and a benchmark report, giving limited auditability. Points are deducted because the benchmark source data is unavailable, no committed tests cover this skill’s key paths, and linked third-party documentation is not reproduced or independently corroborated; the static cap keeps this at or below 5.
- Preview and generation may send prompts, seed fields, or synthetic data to external models or network services; confirm data flow, compliance, and de-identification before using real personal information.
- Autopilot proceeds through validation, preview, and optional creation; beyond the stated large-dataset confirmation, explicitly confirm network calls, cost, output location, and record count before execution.
- Mainland-China reachability, model-alias configuration, and local dependency availability are not verified; provide a local-model or offline fallback where needed.
- The benchmark report omits its raw evaluation dataset and reproducible artifacts, so it should not be treated as independent execution verification.
What does this skill do, and when should you use it?
This skill is for creating datasets, generating synthetic data, and building data-generation pipelines. It directs an agent to use the Data Designer library to produce a Python configuration builder from the user’s dataset description. It supports Interactive mode and Autopilot mode, with Interactive as the default. The instructions define when to read workflow and reference files and provide rules for retaining columns, handling seed data, and configuring validation and templated fields.
Selects the Interactive or Autopilot workflow from the user’s request; reads the matching workflow and, when needed, the person-sampling or seed-dataset reference; builds a synthetic-dataset configuration with the Data Designer API; writes a Python file in the current directory containing load_config_builder(); adds PEP 723 inline dependency metadata; and provides prescribed responses for a missing CLI, Python-version requirements, and preview network failures.
- A data engineer needs a synthetic dataset generated from a natural-language specification.
- An ML team needs a configurable pipeline for producing training or evaluation data.
- A user wants an agent to make generation decisions interactively or autonomously.
- A developer needs sampler, validation, Jinja2 expression, or LLM-judge columns configured correctly.
What are this skill's strengths and limitations?
- Targets synthetic dataset creation and data-generation pipeline construction.
- Offers both Interactive and Autopilot workflows.
- Documents rules for column retention, seed data, person sampling, templating, and LLM-judge score access.
- Defines a concrete Python output interface and dependency format.
- The supplied SKILL.md does not include the contents of either workflow file.
- Execution depends on the Data Designer library and an appropriate runtime environment.
- Preview operations may require network access, but no specific network configuration is documented.
- The `$ARGUMENTS` substitution mechanism requires support from the target agent or light adaptation.
How do you install this skill?
Run npx skills add nvidia/skills --skill data-designer --yes to install the skill in a supported agent environment. The README does not specify the exact local installation path for each agent.
How do you use this skill?
Describe the dataset to the agent, for example: Create a synthetic dataset of customer reviews with ratings and quality scores. Use Autopilot when the user asks the agent to decide or does not want questions; otherwise use the default Interactive mode.