Synthetic Data
Artificially generated data used for training or evaluation.
In plain language
Synthetic data addresses scarce real data, privacy and labeling cost, and is used for instruction tuning, code tests, dialogue simulation and edge-case coverage. Risks include amplified distribution bias, learning the generator’s errors and evaluation contamination. High-quality synthetic data needs explicit generation rules, deduplication, human spot checks and mixing with real data. Evaluating synthetic data with synthetic data is especially dangerous; keep an independent real evaluation set.
Example
Generate 10,000 simulated support dialogues covering refunds, logistics and complaints, but evaluate on real tickets before launch.
Recommended tools
Related terms
Benchmark
A standardized dataset and metric for comparing model capabilities.
Fine-tuning
Continue training a model on task data to better match a style or task.
Model Collapse
Training models repeatedly on their own generated data, causing diversity loss and error accumulation.
Agent Memory
A mechanism for an agent to store and retrieve user preferences, past tasks and key facts across sessions.