This paper introduces MiniCheck, an efficient fact-checking system for large language models (LLMs) that achieves performance comparable to GPT-4 at a significantly lower cost. The key innovation is the creation of synthetic training data that mimics the complexities of fact-checking tasks. By generating challenging factual errors through structured procedures, the system trains models to verify individual facts and recognize information synthesis across sentences. The authors evaluate their approach on a new benchmark, LLM-AGGREFACT, which aggregates multiple existing datasets for both closed-book and grounded generation settings. Their best model, MiniCheck-FT5 (770M parameters), outperforms existing systems and matches GPT-4 accuracy while being 400 times cheaper. The system is efficient, capable of handling multiple facts in a single response, and does not require separate claim decomposition. The paper also shows that claim decomposition is not necessary for high performance, and that decontextualization does not improve performance on their benchmark. Overall, the work demonstrates that synthetic data can be effectively used to train efficient fact-checking models for LLMs.This paper introduces MiniCheck, an efficient fact-checking system for large language models (LLMs) that achieves performance comparable to GPT-4 at a significantly lower cost. The key innovation is the creation of synthetic training data that mimics the complexities of fact-checking tasks. By generating challenging factual errors through structured procedures, the system trains models to verify individual facts and recognize information synthesis across sentences. The authors evaluate their approach on a new benchmark, LLM-AGGREFACT, which aggregates multiple existing datasets for both closed-book and grounded generation settings. Their best model, MiniCheck-FT5 (770M parameters), outperforms existing systems and matches GPT-4 accuracy while being 400 times cheaper. The system is efficient, capable of handling multiple facts in a single response, and does not require separate claim decomposition. The paper also shows that claim decomposition is not necessary for high performance, and that decontextualization does not improve performance on their benchmark. Overall, the work demonstrates that synthetic data can be effectively used to train efficient fact-checking models for LLMs.