Synthetic Data

Definition

Training data generated by AI models rather than collected from humans. Synthetic data generation ranges from simple rephrasing to complex agentic traces with tool calls, multi-step reasoning, and domain-specific structure.

Why Synthetic Data

  1. Scale: Human annotation is expensive and slow. Synthetic generation can produce millions of examples cheaply.
  2. Specialization: Domain-specific data can be generated by fine-tuning a generator model — the generator knows domain rules that general annotators don’t.
  3. Control: Synthetic data can be generated to cover edge cases, adversarial examples, and rare scenarios underrepresented in organic data.
  4. Privacy: No sensitive user data required.

The Generator-Student Pipeline

Frontier labs now routinely use fine-tuned specialist models as generators:

  • Fine-tune a generator on domain expertise
  • Generator produces training data for a student model
  • Student is evaluated; failures feed back to improve generator

This is how Fireworks AI supports Cursor’s Composer-2 RL training. See Fine-Tuning as a Service.

Agentic Data Quality Problem

From Vintage Data: Early agentic traces had ~30% faulty tool calls (malformed JSON). Getting reliable agentic behavior requires high-quality agentic traces — harder to generate than conversational data.

Limitations

  • Distribution shift: Synthetic data can diverge from real-world distribution in subtle ways
  • Error amplification: If the generator model has systematic errors, the student inherits them
  • Formal rule accuracy: Transformers are good at learning formal rules from synthetic data — but only if the synthetic data has correct rules