The Essential RAG Book
Synthetic Data Generation
## 6. Limitations. Synthetic datasets risk encoding the biases of their source LLM.
TL;DR
## 6. Limitations. Synthetic datasets risk encoding the biases of their source LLM.
Key Takeaways
- Moreover, if generated questions too closely mirror the model's pretraining distribution, they may inflate performance metrics.
6. Limitations. Synthetic datasets risk encoding the biases of their source LLM.
Moreover, if generated questions too closely mirror the model's pretraining distribution, they may inflate performance metrics. Diverse sampling and human spot-checks remain critical. When to use: synthetic data generation is most useful for early-stage RAG prototyping, retriever evaluation at scale, and reinforcement of factual grounding during continuous model improvement.
