Synthetic data
Synthetic data is training or test data generated by software, often by another AI model, rather than collected from real-world sources. It is used to fill gaps, add examples cheaply, or avoid exposing real personal data.
By Divya Sharma · Checked Nov 16, 2026 · How we review
Collecting and labelling real data is slow and costly, and some data is private or rare. Synthetic data replaces or adds to it. For language models, a strong model is prompted to write questions, instructions, answers, dialogues or code, and the results are filtered and used for fine-tuning or evaluation.
A well-known example
Self-Instruct (Wang et al., 2022) showed the pattern. A model generates new instructions with inputs and outputs, invalid or near-duplicate items are filtered out, and the original model is fine-tuned on what remains. The authors release the resulting synthetic dataset. Many later instruction datasets follow similar steps, and knowledge distillation often relies on a larger teacher model generating data for a smaller student.
Uses
Common uses include instruction and chat data, rare edge cases, tool-use traces for agents, test sets for LLM evaluation, and stand-in records that look like customer data without being real people. It also shows up in vision, where rendered scenes provide labelled images.
Risks
Synthetic data inherits the generating model's errors, biases and blind spots, and filtering only catches some of them. Training repeatedly on model output with little fresh real data can narrow what a model produces. Generated examples tend to be less varied than real ones. The terms of service of some model providers restrict using their outputs to train competing models, so check them. Keep a real, human-checked golden dataset to measure whether synthetic data actually helped. See also AI training data.
Related terms
- AI training data
- AI training data is the collection of examples a model learns from, such as web text, code, images, transcripts or labelled pairs. Its size, quality, mix and licensing shape what the model can do and where it fails.
- Fine-tuning
- Fine-tuning is further training of a pretrained model on a smaller, task-specific dataset, so the model adopts a particular skill, format, tone or domain without being trained from scratch.
- Knowledge distillation
- Knowledge distillation is a training method in which a smaller student model learns to imitate a larger teacher model's outputs, producing a distilled model that keeps much of the teacher's skill at lower cost.
- Golden dataset
- A golden dataset is a fixed, carefully checked set of example inputs paired with the answers or behaviour you expect, used as the reference when you evaluate or regression-test an AI system.
- LLM evaluation
- LLM evaluation, often called evals, is measuring how well a model or AI application performs a task by running it on a set of test cases and scoring the outputs against defined criteria, by code, by humans or by another model.