LongBench

A bilingual (English + Chinese) benchmark for long-context understanding, with 21 tasks across document QA, summarization, few-shot learning, synthetic reasoning, and code completion. 4,750 test samples averaging 5k–15k tokens, scored with a fully automated evaluation.