When would you use synthetic data to train or test a model, and what are the risks?
Prompting, RAG or fine-tuning · Pro 1
Short answer
When real labelled data is scarce, slow or private: to grow rare classes, cover edge cases, translate into low-resource languages, or label real inputs with a strong model. The risks are errors in generated data, unrealistic language, bias copied from the generator, and model collapse if it replaces real data. Always filter it and test on real data.
Deeper answer
The best value often comes from keeping inputs real and letting a strong model produce the labels, because the data still looks like what users send. Paraphrasing real rare examples is another safe win.
Verification is the heart of it: run generated code or SQL against tests, use an LLM judge with written rules, filter duplicates and length outliers, and read a random sample by hand. Keep real data in the training mix, never test on synthetic data, and record which model generated which data, because provider terms may forbid training competing models on their outputs.
What a strong answer covers
- Use it when real data is scarce, slow or private
- Names generation methods such as labelling, paraphrasing or translation
- Filters and verifies the generated data
- Names risks such as errors, bias or model collapse
- Tests on real, unseen data
Answer it out loud in the mock interview about 1.5 min
Follow-ups to expect
- How would you check the quality of 20,000 model-generated labels?
- Why should the final test set be real data?
Common mistakes
- Testing the model on synthetic data
- Skipping manual review of a sample
- Ignoring the generator's terms of use
Learn it
Synthetic data and distillation: training with model-made data
- How would you build a dataset to fine-tune an LLM for a business task?
- Your fine-tuned model is no better than the base model. How do you debug it?
- Models now have very large context windows. Why not put all the documents in the prompt instead of using RAG?
- You have three LoRA adapters for three tasks on the same base model. Would you merge them, serve them separately, or train one multi-task model?
- How do you decide which model size to use for a feature?
- How would you choose a base model to continue training or fine-tune for a specialised domain or language?
- What is prompt caching, and when does it help?
- When would you use prompting, RAG or fine-tuning?
- How would you build an AI assistant that works well in several Indian languages?
- What is model distillation, and when would you use it?
- When is fine-tuning worth the effort?
- A reasoning model is more accurate on your task but five times the cost and much slower. How do you decide whether to use it?