Instruction-following Evaluation through Verbalizer Manipulation

Shiyang Li, Jun Yan, Hai Wang +4

While instruction-tuned models have shown remarkable success in various natural language processing tasks, accurately evaluating their ability to follow instructions remains challe…