Self-Instruct (2024年)（セルフインストラクト）

概要

Self-Instruct は2024年成熟・Industry-leading LLM self-bootstrap instruction generation 2024年Wang+Mishra+Khashabi Yizhong Wang UW Self-Instruct synthetic instruction generation position確立。Industry-leading 175 seed → 52K instructions + Alpaca + Vicuna basis + GPT-3 self-generated + 2022年12月 paper → 2024 LLM training standard。

主な特徴・仕組み

• Industry-leading Yizhong Wang + Yeganeh Kordi + Swaroop Mishra + Alisa Liu + Noah A. Smith + Daniel Khashabi + Hannaneh Hajishirzi (UW + AI2) 2022年12月 paper → 2024 mature
• Industry-leading Self-Instruct synthetic instruction generation signature
• Industry-leading LLM self-bootstrap instruction signature (model generates own training data)
• Industry-leading 175 seed → 52K instructions signature (350x expansion)
• Industry-leading Alpaca (Stanford) + Vicuna basis signature
• Industry-leading GPT-3 + ChatGPT self-generated training data signature
• Industry-leading instruction-following fine-tuning pioneer signature
• Industry-leading 4-step pipeline signature (instruction + input + output + filtering)
• Industry-leading 2022年12月 ACL 2023 paper "Self-Instruct Aligning Language Model with Self Generated Instructions"
• Industry-leading 2024 instruction tuning ecosystem foundation
• Industry-leading MIT license signature

まとめ

Self-Instruct = 2024年成熟Yizhong Wang UW + AI2発表のpremium synthetic instruction generation。Industry-leading LLM self-bootstrap instruction model generates own training data + 175 seed → 52K instructions 350x expansion + Alpaca Stanford + Vicuna basis + GPT-3 + ChatGPT self-generated training data + instruction-following fine-tuning pioneer + 4-step pipeline instruction + input + output + filtering + 2022年12月 ACL 2023 paper "Self-Instruct Aligning Language Model with Self Generated Instructions" + 2024 instruction tuning ecosystem foundation + MIT license + Yizhong Wang + Yeganeh Kordi + Swaroop Mishra + Alisa Liu + Noah A. Smith + Daniel Khashabi + Hannaneh Hajishirzi UW + AI2 2022→2024 mature position確立。