Synthetic Data

Generate synthetic datasets — from a schema, from a real source, or via LLM-driven persona generation. Includes utilities for PII replacement, real-to-synth transformation, and quality/privacy evaluation.

Skills

• tools-reference — Reference card of recommended OSS tooling (SDV, Synthcity, Faker, Mimesis, DataSynthesizer, ydata-synthetic, Gretel, time-series and LLM options).
• setup-workspace — Initialize a workspace folder (inputs/, outputs/, reports/, configs/) and a requirements.txt.
• tabular-from-schema — Generate tabular data from a JSON schema (Faker/Mimesis + numpy distributions).
• tabular-from-real — Fit SDV (GaussianCopula, CTGAN, TVAE) on a real CSV and sample synthetic rows preserving marginals + correlations.
• replace-pii — Swap PII columns in a real dataset for realistic Faker values, with deterministic mapping for referential integrity.
• text-records-llm — Generate synthetic text records (tickets, reviews, notes) via the Claude CLI with persona/style controls and dedup.
• real-to-synth-llm — LLM-driven transformation of real records into synthetic counterparts that preserve semantic structure but change all specifics.
• evaluate-quality — Fidelity, utility, and privacy diagnostics (SDMetrics for tabular; embedding-based leakage and n-gram diversity for text).

Installation

claude plugins install synthetic-data@danielrosehill

Or, scoped to a single project:

claude plugins install synthetic-data@danielrosehill --scope project

License

MIT