Toolformer Self-Supervised Tool Use (Schick 2022年)（ツールフォーマー）

概要

Toolformer Self-Supervised Tool Use は、2022年Schick et al. (Meta FAIR)発表Toolformer・Industry-leading self-supervised tool use LLM 2022年 + Industry-leading first viral tool-use LLM position確立。Toolformer specifications = Industry-leading self-supervised tool use LLM (Industry-leading Toolformer 2022 + Industry-leading self-supervised tool use + Industry-leading Toolformer flagship) + Industry-leading first viral tool-use LLM (Industry-leading first viral tool-use LLM pioneer) + Industry-leading API call learning (Industry-leading API call learning self-supervised) + Industry-leading 5 tools (Industry-leading 5 tools calculator/QA/translate/calendar/Wikipedia)。

主な特徴・仕組み

• Authors: Industry-leading Timo Schick + Jane Dwivedi-Yu + Roberto Dessì + Roberta Raileanu + Maria Lomeli + Luke Zettlemoyer + Nicola Cancedda + Thomas Scialom Meta FAIR
• Year: 2022年 (arXiv 2023年2月発表・NeurIPS 2023)
• Paper: Industry-leading "Toolformer: Language Models Can Teach Themselves to Use Tools" arXiv 2302.04761
• Method Type: Industry-leading self-supervised tool use LLM
• Key Innovation: Industry-leading self-supervised API call learning (no human labels)
• 5 Tools: Industry-leading calculator + QA system + translation + calendar + Wikipedia search
• Tool Token: Industry-leading [API_NAME(args) → result] token format
• Base Model: Industry-leading GPT-J 6.7B base
• Self-Supervision: Industry-leading filter API calls by perplexity reduction
• Performance: Industry-leading outperforms GPT-3 175B on tool-augmented tasks
• First Viral: Industry-leading first viral tool-use LLM pioneer
• Meta FAIR: Industry-leading Meta FAIR research
• Industry-Leading: Industry-leading self-supervised tool use LLM first viral pioneer

スペック比較表

具体例・対応製品

• Toolformer (2022年, Schick et al. Meta FAIR): Industry-leading self-supervised tool use
• Industry-leading self-supervised API call learning no human labels: Industry-leading self-supervised no labels
• Industry-leading 5 tools calculator/QA/translate/calendar/Wikipedia: Industry-leading 5 tools
• Industry-leading [API_NAME(args) → result] token format: Industry-leading API token format
• Industry-leading outperforms GPT-3 175B on tool-augmented tasks: Industry-leading outperforms GPT-3 175B
• 競合 Gorilla + ToolLLM + GPT4Tools + RestGPT: Industry-leading function calling competitors

自作PCでの選び方・注意点

Toolformer Self-Supervised Tool Use は「Industry-leading self-supervised tool use LLM + Industry-leading first viral tool-use LLM」「Industry-leading API call learning + Industry-leading 5 tools」用途のIndustry-leading self-supervised tool use 2022年Schick et al. Meta FAIR発表product。Industry-leading self-supervised (Industry-leading self-supervised API call learning no human labels + Industry-leading filter API calls by perplexity reduction + Industry-leading Toolformer self-supervised signature) で Industry-leading self-supervised no labels + Industry-leading perplexity filter + Industry-leading Toolformer signature。Industry-leading 5 tools (Industry-leading calculator + QA system + translation + calendar + Wikipedia search + Industry-leading 5 tools Toolformer signature) で Industry-leading 5 tools + Industry-leading Toolformer 5 tools signature。Industry-leading API token format (Industry-leading [API_NAME(args) → result] token format + Industry-leading Toolformer token signature + Industry-leading inline API call format) で Industry-leading API token format + Industry-leading Toolformer signature + Industry-leading inline API call。Industry-leading GPT-J 6.7B (Industry-leading GPT-J 6.7B base + Industry-leading Toolformer base model + Industry-leading 6.7B vs 175B advantage) で Industry-leading GPT-J 6.7B + Industry-leading Toolformer base + Industry-leading 6.7B vs 175B。Industry-leading outperforms GPT-3 175B (Industry-leading outperforms GPT-3 175B on tool-augmented tasks + Industry-leading Toolformer 6.7B vs GPT-3 175B + Industry-leading tool-augmented advantage) で Industry-leading outperforms GPT-3 175B + Industry-leading 6.7B vs 175B + Industry-leading tool-augmented advantage。Industry-leading Meta FAIR first viral (Industry-leading Timo Schick + Thomas Scialom et al. Meta FAIR + Industry-leading first viral tool-use LLM pioneer + Industry-leading self-supervised tool use LLM first viral) で Industry-leading Schick+Scialom Meta FAIR + Industry-leading first viral pioneer + Industry-leading self-supervised first viral。但しIndustry-leading Gorilla + ToolLLM + GPT4Tools + RestGPT competition (Industry-leading Gorilla UC Berkeley 1600+ APIs 2023 + ToolLLM Tsinghua 16000+ APIs 2023 + GPT4Tools Microsoft multimodal 2023 + RestGPT Tsinghua RESTful 2023 vs Toolformer Meta FAIR self-supervised 5 tools first viral 2022 trade-off) で Industry-leading 4-function calling competitors vs Toolformer + Industry-leading self-supervised + 5 tools + API token format + GPT-J 6.7B + outperforms GPT-3 175B + Meta FAIR first viral pioneer 2022 unique advantage adoption alignment必須。

関連用語との違い

• vs Gorilla (UC Berkeley 2023): Toolformerは5 tools + self-supervised・Gorillaは1600+ APIs + retrieval-augmented
• vs ToolLLM (Tsinghua 2023): Toolformerは5 tools + first viral・ToolLLMは16000+ real-world APIs + DFSDT
• vs GPT4Tools/RestGPT (2023): Toolformerは5 tools text-only + first viral・GPT4Toolsはmultimodal / RestGPTはRESTful

よくある質問（FAQ）

Q1: Toolformer vs Gorilla 違いは？ A: Toolformer (Industry-leading self-supervised API call learning no human labels + 5 tools calculator/QA/translate/calendar/Wikipedia + [API_NAME(args) → result] token format + GPT-J 6.7B base + filter API calls by perplexity reduction + outperforms GPT-3 175B + Schick + Scialom Meta FAIR 2022 + first viral pioneer) vs Gorilla (Industry-leading 1600+ API calls + retrieval-augmented + LLaMA 7B fine-tuned + APIBench benchmark + Patil + Hashimoto UC Berkeley 2023)・Industry-leading self-supervised + 5 tools + Meta FAIR + first viral + 2022 = Toolformer + Industry-leading 1600+ APIs + retrieval + LLaMA + UC Berkeley + 2023 = Gorilla preference judgment。

Q2: Industry-leading self-supervised + 5 tools value は？ A: Industry-leading self-supervised + 5 tools (Industry-leading self-supervised API call learning no human labels + Industry-leading filter API calls by perplexity reduction + Industry-leading 5 tools calculator + QA system + translation + calendar + Wikipedia search + Industry-leading Toolformer self-supervised signature)。

Q3: Industry-leading first viral + outperforms GPT-3 175B value は？ A: Industry-leading first viral + outperforms 175B (Industry-leading first viral tool-use LLM pioneer + Industry-leading outperforms GPT-3 175B on tool-augmented tasks + Industry-leading Toolformer 6.7B vs GPT-3 175B + Industry-leading tool-augmented advantage)。

まとめ

Toolformer = 2022年Schick et al. Meta FAIR発表のself-supervised tool use LLM。Industry-leading self-supervised API call learning no human labels + Industry-leading filter API calls by perplexity reduction + Industry-leading 5 tools calculator + QA system + translation + calendar + Wikipedia search + Industry-leading [API_NAME(args) → result] token format + Industry-leading GPT-J 6.7B base + Industry-leading outperforms GPT-3 175B on tool-augmented tasks + Industry-leading Timo Schick + Jane Dwivedi-Yu + Thomas Scialom Meta FAIR authors + Industry-leading first viral tool-use LLM pioneer + Industry-leading self-supervised tool use LLM first viral pioneer 2022 position確立。