Synthetic Data Generation

Purpose: Build training and eval datasets quickly without violating user privacy, while controlling for quality, diversity, and contamination.

When to Use This Skill

• Bootstrapping eval sets when you have no real labeled data
• Augmenting fine-tuning data for under-represented intents / slices
• Generating adversarial test cases (combine with ai-safety-and-red-teaming)
• Creating QA pairs over a private RAG corpus
• Privacy-safe substitutes for production data in dev/test

When NOT to Use

• If real, well-labeled data exists at sufficient scale -- use it
• For final benchmark numbers in regulated decisions -- use human-labeled holdouts