Open AkihikoWatanabe opened 11 months ago
seed instructionとdemonstrationに基づいて、異なるスタイルのinstructionを自動生成し、自動生成したinstructionをとinferenceしたいexampleで条件づけてランキングし、良質なものを選択。選択したinstructionでinferenceを実施する。
既存手法よりも高い性能を達成している。特にexampleごとにinstructionを選択する手法の中で最もgainが高い。これは、提案手法がinstructionの選択にtrained modelを利用しているためであると考えられる。
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)