WildIFEval: Instruction Following in the Wild

WildIFEval is a newly released dataset comprising 7,000 real user instructions that feature diverse multi-constraint conditions, aimed at enhancing instruction-following capabilities in large language models (LLMs). The dataset categorizes constraints into eight high-level classes and has been used to benchmark various LLMs, revealing significant performance gaps between small and large models, as well as highlighting areas for improvement in handling complex instructions. This resource is crucial for practitioners focusing on developing LLMs that can effectively interpret and execute nuanced user commands in real-world applications.