Synthetic data is artificially generated data that resembles real user data without being based on real-life records. In AI training and fine-tuning, it can supply useful examples when suitable real data is limited, costly, or difficult to prepare. It can also help generate or transform test data rapidly, reducing the effort required for test-data management. section 12.2.1 section 1.1.1
Use synthetic data with clear limits:
- Expect generation costs: services such as Gretel.ai, Mostly AI, and Tonic AI provide data-generation tools, but they are paid tools. section 12.2.1
- Retain human judgment and verification because purely synthetic training data can reduce response quality by missing the variance and randomness of real data. section 12.2.1
- Treat dataset size separately from usefulness: Alpaca contains 52,000 synthetic instruction examples, yet remains a useful dataset alongside much larger collections. section 12.2.1
Instruction-and-output records
A synthetic dataset can pair a generated instruction such as “What is the class declaration for BrandingServiceIT?” with a formatted Java answer. Similar records can be used as examples for training or fine-tuning. section 12.2.1
