CIF-Bench: A Chinese Instruction-Following Benchmark for Evaluating the Generalizability of Large Language Models