| name | tools-reference |
| description | Reference card of open-source synthetic data tools, when to use each, install commands, and design patterns. |
Tools Reference
A curated reference of open-source and cloud synthetic data tools. Use this skill when the user asks which tool to pick for their use case, or wants to understand the landscape.
Tabular data
SDV (Synthetic Data Vault)
What it does: Learns probability distribution from real data, then samples synthetic rows that preserve marginals, correlations, and constraints.
When to use:
- Real tabular data → synthetic tabular data (preserving statistical properties)
- Fast iteration on schema-less synthesis
- Single-table or multi-table (relational) datasets
Pros: Mature, multi-table support, handles mixed types, row/column constraints, reversible transformations.
Cons: Slower for very high-dimensional data; assumes stationarity.
Key models:
GaussianCopulaSynthesizer — fast, general, works well for continuous/categorical mixCTGANSynthesizer — neural-network-based, high fidelity on mixed types, slower to fitTVAESynthesizer — variational autoencoder, good balance of speed/fidelityHMASynthesizer — multi-table, learns foreign-key relationships
Install: pip install sdv
Quality eval: pip install sdmetrics
Synthcity
What it does: Plugin-based generative synthesis with support for differential privacy, fairness constraints, and time-series.
When to use:
- Privacy-sensitive tabular synthesis with formal privacy budgets
- Time-series generation
- Fairness/bias constraints on synthetic output
Pros: DP guarantees, plugin architecture, time-series module.
Cons: Steeper learning curve; fewer pre-packaged models than SDV.
Install: pip install synthcity
DataSynthesizer
What it does: Differentially-private tabular synthesis using Gaussian copula or uniform sampling.
When to use:
- Data that must satisfy formal differential-privacy (DP) constraints
- Compliance-heavy industries (HIPAA, GDPR)
- Schema is known or easily inferred
Pros: Formal privacy guarantees, simple API, fast.
Cons: Limited to DP copula approach; less flexible than SDV.
Install: pip install DataSynthesizer
ydata-synthetic
What it does: GAN-based tabular and time-series synthesis.
When to use:
- High-fidelity tabular synthesis where neural approaches needed
- Time-series with adversarial training
Pros: Strong empirical fidelity; time-series support.
Cons: Training can be unstable; hyperparameter-sensitive.
Install: pip install ydata-synthetic
Field-level (record-wise) generation
Faker
What it does: Generates realistic fake values for common fields (names, emails, phone numbers, addresses, dates, credit cards, SSNs, IP addresses, etc.) with locale support.
When to use:
- Schema-based generation (you define columns, Faker fills values)
- PII replacement (swap real names/emails with fake ones)
- Creating realistic test data quickly
- Localized personas (de_DE, fr_FR, ja_JP, etc.)
Pros: Simple, fast, extremely customizable, 50+ locales.
Cons: No correlation learning; purely random per-field.
Install: pip install faker
Example:
from faker import Faker
fake = Faker('de_DE')
print(fake.name(), fake.email(), fake.phone_number())
Mimesis
What it does: Similar to Faker — generates localized fake data with additional structure (e.g. geography, person relationships).
When to use:
- Alternative to Faker with slightly different provider library
- Data generation with geographic/organizational context
Pros: Compact, fast, good geographic support.
Cons: Smaller ecosystem than Faker.
Install: pip install mimesis
Time-series
Chronos (Amazon)
What it does: LLM-based probabilistic time-series forecasting and generation.
When to use:
- Synthetic time-series that extrapolate real historical patterns
- Pre-trained foundation model approach (zero-shot)
Pros: Pre-trained, handles multiple series lengths.
Cons: Requires external service or local inference.
Install: Via Hugging Face, huggingface_hub
TimeGAN
What it does: Generative adversarial network for time-series synthesis.
When to use:
- Synthetic sequences that preserve temporal dependencies
- Long sequences with complex dynamics
Pros: Good temporal fidelity.
Cons: Training instability; requires careful hypertuning.
Install: Via research implementations (e.g. GitHub repos)
Unstructured/Text
Claude API (LLM-driven)
What it does: Use Claude to generate or transform text records based on prompts describing persona, tone, intent, schema.
When to use:
- Synthetic text (support tickets, reviews, medical notes, chat logs)
- Real-to-synthetic transformation (change specifics, preserve semantics)
- High-quality, semantically coherent output preferred over speed
Pros: Flexible, human-readable output, custom logic via prompts.
Cons: API costs; slower than generative models; requires API key.
Install: pip install anthropic
GPT / OpenAI API
What it does: Similar to Claude but via OpenAI models.
When to use:
- Text generation when GPT models preferred
- Batch operations via OpenAI batch API
Install: pip install openai
Cloud / Fully-managed
Gretel CLI
What it does: Client for Gretel's cloud-based synthetic data platform. Uploads data, trains models in the cloud, downloads synthetic data.
When to use:
- User prefers cloud-hosted synthesis (no local infra)
- Need advanced options (differential privacy, data transforms) without local setup
- Data governance / audit trail required
Pros: Managed service, no local compute, audit trail.
Cons: Costs; data leaves local network.
Install: pip install gretel-client
Choosing a tool
| Use case | Recommended | Alternatives |
|---|
| Real tabular → synthetic | SDV (GaussianCopula) | Synthcity, ydata-synthetic |
| Mixed-type tabular, high fidelity | SDV (CTGAN) | ydata-synthetic (GAN), Gretel |
| Privacy-sensitive tabular | Synthcity, DataSynthesizer | SDV + manual DP |
| Multi-table relational | SDV (HMA) | Synthcity with custom |
| Schema-based tabular (fake) | Faker + Mimesis | custom generation |
| PII replacement on real data | Faker + deterministic mapping | ydata-synthetic (PII aware) |
| Synthetic text records | Claude/GPT API | ydata-synthetic (text models) |
| Real text → synthetic (paraphrase) | Claude/GPT API with prompts | fine-tuned LLMs |
| Time-series forecasting | Chronos | SDV (multivariate time) |
| Time-series GAN-based | TimeGAN, ydata-synthetic | SDV (univariate) |
Next: Choose your use case, then refer to the appropriate skill (tabular-from-schema, tabular-from-real, text-records-llm, etc.) for step-by-step execution.
