Synthetic Data
Synthetic data is training data created by a model (usually a stronger model or a specialized generator), rather than scraped from the real world. When real data is scarce (medical, rare languages, niche domains), too expensive (human labeling), or too sensitive (user conversations, clinical records), synthetic data is the common way to bypass those limits.
Main uses
1. Supervised fine-tuning (SFT) augmentation
- Use GPT-4 / Claude to generate diverse instruction-response pairs
- Phi (Microsoft), WizardLM, Self-Instruct all go this route
- Key insight: diversity beats quantity; homogeneous data makes the model regress
2. RL reward / verification data
- Generate large volumes of "ground-truth-answer" problems (math, code, reading comprehension)
- No human labeling needed; verifier auto-grades (see
rlvr.md)
3. Preference data
- A strong model ranks multiple candidates for the same prompt
- Replaces expensive human preference annotation in RLHF
- UltraFeedback / HH-RLHF both use this trick
4. The extreme case: 100% synthetic training
TypeSafe AI Jev is the publicly-documented industrial-grade model trained entirely on synthetic data: no real user data in the training set.
- A program generates every schema → decision pair
- RLCD optimizes calibration directly on synthetic data
- Smaller risk surface (no privacy / copyright / bias concerns)
Risks and failure modes
Model Collapse
- Training AI on recursively generated data causes exponential decay to meaningless output
- First systematically proven in a 2024 Nature paper, "AI models collapse when trained on recursively generated data"
- Mitigation: mix a fraction of real data per generation; use a weaker model to supervise
Variable quality
- Generator errors / biases get inherited by downstream models
- Requires verifier / filtering / human spot-checks
Homogenization
- Generators trend toward "modal" output (default style, average length)
- Low diversity makes downstream models dumber
- Mitigation: raise temperature + multi-seed + multi-prompt paraphrasing
Legal and ethical
- Copyright: whether synthetic counts as "use of the original" is disputed
- Privacy: if the generator saw user data, the output may contain PII
Evaluation methods
- Human eval: spot-check with human raters
- Downstream benchmark: watch for regression on standard test sets
- Diversity metrics: n-gram entropy, embedding distribution
- Backtest: after training, evaluate on real distribution data and measure the drop
Tooling
- GPT-4 / Claude API: strongest but most expensive, best for small high-quality batches
- Open-source LLM + custom pipeline: best price/quality, needs self-built quality filter
- Specialized generators: Cosmopedia (educational), Synthia (instruction)
- Curriculum learning: generate from easy to hard to avoid models getting stuck on simple patterns