Self-Instruct

by University of Washington · open-source · Last verified 2026-03-17

Self-Instruct is the foundational instruction-tuning dataset and methodology introduced by Wang et al. (2022), where 175 human-written seed tasks are iteratively expanded into 52,000 instruction-input-output triplets using GPT-3 as the generator. It established the paradigm of bootstrapping instruction data from existing LLMs and directly inspired Alpaca, WizardLM, and most subsequent synthetic alignment datasets.