Instruction-following Evaluation through Verbalizer Manipulation

arXiv:2307.10558v2 Announce Type: replace
Abstract: While instruction-tuned models have shown remarkable success in various natural language processing tasks, accurately evaluating their ability to follow instructions remains challenging. Existing benchmarks primarily focus on common instructions that align well with what the model learned during training. However, proficiency in responding to these instructions does not necessarily imply strong ability in instruction following. In this paper, we propose a novel instruction-following evaluation protocol called verbalizer manipulation. It instructs the model to verbalize the …

abstract arxiv benchmarks cs.cl evaluation focus however imply instruction-tuned language language processing manipulation natural natural language natural language processing processing success tasks through training type