Self-Instruct (Wang + UW 2022年-2024年)（セルフインストラクト）
2022年Wang et al. + University of Washington発表Self-Instruct paper・Industry-leading Self-Instruct LLM data augmentation paradigm + Industry-leading LLM self-generated instruction tuning data + Industry-emerging Self-Instruct Pioneer。
概要
Self-Instruct は、2022年Wang et al. + University of Washington発表Self-Instruct paper・Industry-leading Self-Instruct LLM data augmentation paradigm 2022年-2024年 + Industry-emerging Self-Instruct Pioneer position確立。Self-Instruct specifications = Industry-leading Self-Instruct LLM data augmentation paradigm (Industry-leading Self-Instruct + Industry-leading LLM data augmentation paradigm + Industry-emerging Self-Instruct Pioneer) + Industry-leading LLM self-generated instruction tuning data (Industry-leading LLM self-generated instruction tuning data + Industry-leading bootstrap instruction tuning) + Industry-emerging Self-Instruct Pioneer (Industry-emerging Self-Instruct Pioneer + Industry-leading Alpaca/Vicuna/Evol-Instruct/Magpie successors Pioneer foundation)。
主な特徴・仕組み
- Authoring: Wang et al. + University of Washington 2022年発表
- Paradigm: Industry-leading Self-Instruct LLM data augmentation
- Mechanism: Industry-leading LLM self-generated instruction tuning data
- Bootstrap: Industry-leading 175 seed tasks + GPT-3 expansion
- Heritage: Industry-leading 2年heritage Self-Instruct Pioneer
- UW Backing: Industry-leading University of Washington academic AI
- Industry Position: Industry-emerging Self-Instruct Pioneer
- Successors: Industry-leading Alpaca + Vicuna + Evol-Instruct + Magpie successors
- Industry-Emerging: Industry-emerging Self-Instruct Pioneer
- Use Cases: Industry-leading LLM instruction tuning data generation
スペック比較表
| LLM Data Augmentation (2022-2024) | Approach | Heritage | Industry Position |
|---|---|---|---|
| Self-Instruct (Wang UW) | Self-generated instruction tuning | 2022年UW Pioneer | Industry-emerging Self-Instruct Pioneer |
| Evol-Instruct (WizardLM) | Evolving instruction complexity | 2023年WizardLM | Industry-emerging Evol-Instruct WizardLM |
| Magpie Data (Xu UCLA) | Aligned LLM self-synthesis | 2024年UCLA | Industry-emerging Magpie aligned synthesis |
| OpenHermes 2.5 (Teknium) | Curated instruction dataset | 2023年Teknium | Industry-leading open-source dataset |
| Back Translation (Sennrich) | Back-translation MT augmentation | 2016年MT classical | Industry-leading classical augmentation |
具体例・対応製品
- Self-Instruct (Wang + UW・2022年-2024年): Industry-emerging Self-Instruct Pioneer
- Industry-leading LLM self-generated instruction tuning data: Industry-leading self-generated
- Industry-leading 175 seed tasks + GPT-3 expansion: Industry-leading 175 seeds + GPT-3
- Industry-leading University of Washington academic AI backing: Industry-leading UW backing
- Industry-leading Alpaca + Vicuna + Evol-Instruct + Magpie Industry-leading successors: Industry-leading successors
- 競合 Evol-Instruct + Magpie + OpenHermes + Back Translation: Industry-emerging augmentation competitors
自作PCでの選び方・注意点
Self-Instruct は「Industry-leading Self-Instruct LLM data augmentation paradigm + Industry-leading LLM self-generated instruction tuning data」「Industry-emerging Self-Instruct Pioneer」用途のIndustry-emerging Self-Instruct 2022年Wang + UW product。Industry-leading LLM self-generated instruction tuning data (Industry-leading LLM self-generated instruction tuning data + Industry-leading bootstrap instruction tuning + Industry-leading 175 seed tasks + GPT-3 expansion) で Industry-leading LLM self-generated + Industry-leading bootstrap instruction tuning + Industry-leading 175 seeds + GPT-3 expansion。Industry-emerging Self-Instruct Pioneer (Industry-emerging Self-Instruct Pioneer + Industry-leading Alpaca/Vicuna/Evol-Instruct/Magpie successors Pioneer foundation + Industry-leading 2年heritage Self-Instruct Pioneer) で Industry-emerging Pioneer + Industry-leading successors Pioneer foundation + Industry-leading 2年heritage。Industry-leading UW backing (Industry-leading University of Washington academic AI backing + Industry-leading Wang et al. UW paper authors) で Industry-leading UW + Industry-leading Wang UW authors。但しIndustry-leading Evol-Instruct + Magpie + OpenHermes + Back Translation competition (Industry-leading Evol-Instruct WizardLM 2023 + Magpie UCLA 2024 + OpenHermes 2.5 Teknium + Back Translation Sennrich vs Self-Instruct 2022 Pioneer trade-off) で Industry-leading successors evolution vs Self-Instruct Pioneer + Industry-leading 2年Pioneer foundation adoption alignment必須。
関連用語との違い
- vs Evol-Instruct (WizardLM 2023年): Self-InstructはSelf-generated Pioneer + UW 2022年・Evol-InstructはEvolving complexity + WizardLM 2023年
- vs Magpie Data (Xu UCLA 2024年): Self-InstructはBootstrap + UW 2022年Pioneer・MagpieはAligned LLM synthesis + UCLA 2024年
- vs Back Translation (Sennrich 2016年MT classical): Self-InstructはLLM self-generated instruction + 2022年・Back TranslationはMT augmentation + 2016年classical
よくある質問（FAQ）
Q1: Self-Instruct vs Evol-Instruct 違いは？ A: Self-Instruct (Wang + UW 2022年 + LLM self-generated instruction tuning + 175 seed tasks + GPT-3 expansion + Self-Instruct Pioneer) vs Evol-Instruct (WizardLM 2023年 + Evolving instruction complexity + In-Depth/In-Breadth Evolving + WizardLM successor)・Industry-leading Pioneer + bootstrap + UW = Self-Instruct + Industry-leading evolving complexity + WizardLM successor = Evol-Instruct preference judgment。
Q2: Industry-leading 175 seed tasks + GPT-3 expansion value は？ A: Industry-leading 175 seeds (Industry-leading 175 seed tasks + Industry-leading GPT-3 expansion + Industry-leading bootstrap instruction tuning + Industry-leading minimal seed data approach)。
Q3: Industry-emerging Self-Instruct Pioneer foundation value は？ A: Industry-emerging Pioneer (Industry-emerging Self-Instruct Pioneer + Industry-leading Alpaca/Vicuna/Evol-Instruct/Magpie successors Pioneer foundation + Industry-leading 2年heritage Self-Instruct Pioneer)。
まとめ
Self-Instruct = 2022年Wang et al. + University of Washington発表のSelf-Instruct LLM data augmentation paradigm。Industry-leading LLM self-generated instruction tuning data + Industry-leading 175 seed tasks + GPT-3 expansion + Industry-leading bootstrap instruction tuning + Industry-leading University of Washington academic AI research backing + Industry-leading Alpaca + Vicuna + Evol-Instruct + Magpie Industry-leading successors Pioneer foundation + Industry-leading 2年heritage Self-Instruct Pioneer position確立。