What is synthetic data generation for fine-tuning?
LLM & AILevel: IntermediateDataset Preparation
What is synthetic data generation for fine-tuning?
Hints
- Which parameters change and which stay frozen?
- Separate memory for weights, gradients, optimizer states and activations.
Answer
When real labeled data is scarce, use a strong LLM to generate training data for a weaker model:
- Seed examples — start with 50–200 human-curated examples
- Prompt a strong model (GPT-4) — "Generate 100 more examples similar to these, with diversity in [X]"
- Quality filter — apply heuristics or use another LLM call to score/filter generated examples
- Human review — sample 5–10% of synthetic data for human quality review
Legal consideration: check the terms of service of the model generating synthetic data. Some providers (OpenAI) prohibit using API outputs to train competing models.
Risks: distribution collapse (synthetic data narrows the distribution); error amplification (if seed examples have biases, synthetic data amplifies them). Always evaluate on real-world held-out data.
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0