CC-Bench

It integrates the multilingual alignment features of the MKQA dataset with the high-quality, human-annotated question-answering pairs from the TyDi QA dataset, utilizing its "unanswerable" labels to distinguish between known and unknown question sets. The benchmark comprises 10,000 cross-domain evaluation samples spanning fields such as sports, history, and basic medicine, and is further categorized into factual and text-based subsets.

It is primarily used to quantify cross-lingual semantic stability, supporting cognitive boundary detection, retrieval-augmented gating optimization, and hallucination mitigation in black-box scenarios.