Synthetic data

Bootstrapping eval sets or formatting examples when real data is scarce.

As a silent substitute for regulated or sensitive real records without review.

LLM writes 200 JSON formatting pairs from seed templates — human spot-checks 10% before fine-tune.