Guides/Synthetic data
Synthetic data and augmentation
Real data is expensive, imbalanced, or too sensitive to use directly — synthetic data fills gaps, but it's a tool with real failure modes, not a free lunch. Match the pattern to the gap, and always audit for privacy leakage and distribution drift before training on synthetic output.
When synthetic data helps — and when it doesn't
Where it helps
- Real data is scarce or heavily imbalanced (rare classes, edge cases)
- Real data is too sensitive to use directly (PII, health, financial records)
- You need adversarial or edge-case examples for stress-testing safety
- You need scale to bootstrap an eval set before real usage data exists
Where it's risky
- Substituting for real-world messiness in the final production eval
- Training successive model generations only on prior synthetic output
- Skipping a privacy audit because the data is “synthetic”
- Using it to hide a real data-collection gap instead of closing it
Patterns and tradeoffs
| Pattern | Core idea | Upside | Watch-outs |
|---|---|---|---|
| Data augmentation | Transform or perturb existing real examples (paraphrase, noise, rotation) | Cheap, preserves the real distribution | Doesn't add genuinely new signal |
| LLM-generated examples | Prompt an LLM to generate labeled training or eval examples | Fast to bootstrap eval sets and rare classes | Can encode the generating model's own biases and blind spots |
| Simulation / programmatic generation | Rule-based or simulated environments generate data and labels | Full control over edge cases and ground truth | Sim-to-real gap; can miss real-world messiness |
| Privacy-preserving synthetic replicas | A generative model trained on real data produces a synthetic replica dataset | Enables sharing and testing without exposing PII | Can still leak signal about real records if not audited |
Watch for model collapse
Training successive model generations on their own (or each other's) synthetic output is a documented failure mode — quality and diversity degrade over generations as errors compound and rare patterns get filtered out. Keep a real-data anchor in every training and eval set, and treat “100% synthetic” pipelines as a research bet, not a default.