ToolLLM 16000+ API Tool LLM (Qin 2023年)（ツールエルエルエム）
2023年Qin et al. (Tsinghua)発表ToolLLM・Industry-leading 16000+ real-world API LLM + Industry-leading DFSDT depth-first search + Industry-leading ToolBench dataset + Industry-leading LLaMA-7B fine-tuned。
概要
ToolLLM 16000+ API Tool LLM は、2023年Qin et al. (Tsinghua)発表ToolLLM・Industry-leading 16000+ real-world API LLM 2023年 + Industry-leading DFSDT depth-first search position確立。ToolLLM specifications = Industry-leading 16000+ real-world API LLM (Industry-leading ToolLLM 2023 + Industry-leading 16000+ real-world API + Industry-leading ToolLLM flagship) + Industry-leading DFSDT depth-first search (Industry-leading DFSDT depth-first search-based decision tree) + Industry-leading ToolBench dataset (Industry-leading ToolBench dataset 16000+ APIs) + Industry-leading LLaMA-7B fine-tuned (Industry-leading LLaMA-7B fine-tuned)。
主な特徴・仕組み
- Authors: Industry-leading Yujia Qin + Shihao Liang + Yining Ye + Kunlun Zhu + Lan Yan + Yaxi Lu + Yankai Lin + Xin Cong + Xiangru Tang + Bill Qian + Sihan Zhao + Lauren Hong + Runchu Tian + Ruobing Xie + Jie Zhou + Mark Gerstein + Dahai Li + Zhiyuan Liu + Maosong Sun Tsinghua + ModelScope + Renmin
- Year: 2023年 (arXiv 2023年7月発表 + ICLR 2024)
- Paper: Industry-leading "ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs" arXiv 2307.16789
- Method Type: Industry-leading 16000+ real-world API LLM
- APIs: Industry-leading 16,464 real-world APIs from RapidAPI Hub 49 categories
- DFSDT: Industry-leading DFSDT (Depth-First Search-based Decision Tree)
- ToolBench: Industry-leading ToolBench dataset 16000+ APIs 126,486 instances
- Base Model: Industry-leading LLaMA-7B fine-tuned (ToolLLaMA)
- ToolEval: Industry-leading ToolEval automatic evaluator
- Performance: Industry-leading ToolLLaMA outperforms GPT-3.5 on ToolEval
- Real-World: Industry-leading real-world APIs (vs synthetic Toolformer)
- Open Source: Industry-leading ToolLLM GitHub OpenBMB/ToolBench
- Industry-Leading: Industry-leading 16000+ real-world API LLM Tsinghua
スペック比較表
| LLM Function Calling (2022-2024) | Year | Author/Org | Key Feature | Industry Position |
|---|---|---|---|---|
| ToolLLM | 2023 | Qin Tsinghua | 16000+ real-world APIs + DFSDT | Industry-leading 16000+ APIs |
| Gorilla | 2023 | Patil UC Berkeley | 1600+ API calls + retrieval-augmented | Industry-leading 1600+ APIs |
| Toolformer | 2022 | Schick Meta FAIR | Self-supervised tool use + 5 tools | Industry-leading first viral |
| GPT4Tools | 2023 | Yang Microsoft | Multimodal tool use + visual tasks | Industry-leading multimodal tools |
| RestGPT | 2023 | Song Tsinghua | RESTful API control + online planner | Industry-leading RESTful API |
具体例・対応製品
- ToolLLM (2023年, Qin et al. Tsinghua): Industry-leading 16000+ real-world API LLM
- Industry-leading 16,464 real-world APIs RapidAPI Hub 49 categories: Industry-leading 16,464 RapidAPI
- Industry-leading DFSDT Depth-First Search-based Decision Tree: Industry-leading DFSDT
- Industry-leading ToolBench dataset 126,486 instances: Industry-leading ToolBench 126,486
- Industry-leading ToolLLaMA outperforms GPT-3.5 on ToolEval: Industry-leading outperforms GPT-3.5
- 競合 Toolformer + Gorilla + GPT4Tools + RestGPT: Industry-leading function calling competitors
自作PCでの選び方・注意点
ToolLLM 16000+ API Tool LLM は「Industry-leading 16000+ real-world API LLM + Industry-leading DFSDT depth-first search」「Industry-leading ToolBench dataset + Industry-leading LLaMA-7B fine-tuned」用途のIndustry-leading 16000+ real-world API 2023年Qin et al. Tsinghua発表product。Industry-leading 16,464 RapidAPI (Industry-leading 16,464 real-world APIs from RapidAPI Hub 49 categories + Industry-leading ToolLLM 16000+ APIs signature + Industry-leading real-world vs synthetic) で Industry-leading 16,464 RapidAPI + Industry-leading ToolLLM 16000+ signature + Industry-leading real-world vs synthetic。Industry-leading DFSDT (Industry-leading DFSDT Depth-First Search-based Decision Tree + Industry-leading ToolLLM DFSDT signature + Industry-leading API selection algorithm) で Industry-leading DFSDT + Industry-leading ToolLLM DFSDT signature + Industry-leading API selection。Industry-leading ToolBench (Industry-leading ToolBench dataset 16000+ APIs 126,486 instances + Industry-leading ToolLLM ToolBench signature + Industry-leading large-scale tool dataset) で Industry-leading ToolBench 126,486 + Industry-leading ToolLLM ToolBench signature + Industry-leading large-scale tool。Industry-leading ToolLLaMA + ToolEval (Industry-leading LLaMA-7B fine-tuned ToolLLaMA + Industry-leading ToolEval automatic evaluator + Industry-leading ToolLLaMA outperforms GPT-3.5 on ToolEval) で Industry-leading ToolLLaMA + Industry-leading ToolEval + Industry-leading outperforms GPT-3.5。Industry-leading Tsinghua + OpenBMB (Industry-leading Yujia Qin + Shihao Liang + Yankai Lin + Zhiyuan Liu + Maosong Sun Tsinghua authors + Industry-leading ToolLLM GitHub OpenBMB/ToolBench + Industry-leading 16000+ real-world API LLM Tsinghua) で Industry-leading Qin+Liang+Lin+Liu+Sun Tsinghua + Industry-leading OpenBMB/ToolBench + Industry-leading Tsinghua 16000+ real-world。但しIndustry-leading Toolformer + Gorilla + GPT4Tools + RestGPT competition (Industry-leading Toolformer Meta FAIR 5 tools self-supervised 2022 + Gorilla UC Berkeley 1600+ APIs retrieval 2023 + GPT4Tools Microsoft multimodal 2023 + RestGPT Tsinghua RESTful 2023 vs ToolLLM Tsinghua 16000+ real-world APIs DFSDT ToolBench 2023 trade-off) で Industry-leading 4-function calling competitors vs ToolLLM + Industry-leading 16,464 real-world APIs RapidAPI 49 categories + DFSDT + ToolBench 126,486 instances + LLaMA-7B fine-tuned ToolLLaMA + ToolEval + outperforms GPT-3.5 + Tsinghua + OpenBMB 2023 unique advantage adoption alignment必須。
関連用語との違い
- vs Gorilla (UC Berkeley 2023): ToolLLMは16000+ real-world APIs + DFSDT + ToolBench・Gorillaは1600+ APIs + retrieval + APIBench
- vs Toolformer (Meta FAIR 2022): ToolLLMは16000+ real-world APIs + DFSDT・Toolformerは5 tools + self-supervised + synthetic
- vs GPT4Tools/RestGPT (2023): ToolLLMは16000+ APIs text-based・GPT4Toolsはmultimodal / RestGPTはRESTful
よくある質問（FAQ）
Q1: ToolLLM vs Gorilla 違いは？ A: ToolLLM (Industry-leading 16,464 real-world APIs from RapidAPI Hub 49 categories + DFSDT Depth-First Search-based Decision Tree + ToolBench dataset 16000+ APIs 126,486 instances + LLaMA-7B fine-tuned ToolLLaMA + ToolEval automatic evaluator + outperforms GPT-3.5 + Qin + Liang + Liu Tsinghua + OpenBMB 2023) vs Gorilla (Industry-leading 1600+ APIs from TorchHub + TensorHub + HuggingFace + retrieval-augmented + LLaMA 7B fine-tuned + APIBench benchmark + hallucination reduction vs GPT-4 + Patil + Gonzalez UC Berkeley 2023)・Industry-leading 16000+ + DFSDT + ToolBench + Tsinghua = ToolLLM + Industry-leading 1600+ + retrieval + APIBench + UC Berkeley = Gorilla preference judgment。
Q2: Industry-leading 16,464 real-world APIs + DFSDT value は？ A: Industry-leading 16,464 + DFSDT (Industry-leading 16,464 real-world APIs from RapidAPI Hub 49 categories + Industry-leading DFSDT Depth-First Search-based Decision Tree + Industry-leading API selection algorithm + Industry-leading ToolLLM signature)。
Q3: Industry-leading ToolBench 126,486 + ToolLLaMA outperforms GPT-3.5 value は？ A: Industry-leading ToolBench + outperforms (Industry-leading ToolBench dataset 16000+ APIs 126,486 instances + Industry-leading LLaMA-7B fine-tuned ToolLLaMA + Industry-leading ToolEval automatic evaluator + Industry-leading ToolLLaMA outperforms GPT-3.5 on ToolEval)。
まとめ
ToolLLM = 2023年Qin et al. Tsinghua + OpenBMB発表の16000+ real-world API LLM。Industry-leading 16,464 real-world APIs from RapidAPI Hub 49 categories + Industry-leading DFSDT Depth-First Search-based Decision Tree + Industry-leading ToolBench dataset 16000+ APIs 126,486 instances + Industry-leading LLaMA-7B fine-tuned ToolLLaMA + Industry-leading ToolEval automatic evaluator + Industry-leading ToolLLaMA outperforms GPT-3.5 on ToolEval + Industry-leading Yujia Qin + Shihao Liang + Yankai Lin + Zhiyuan Liu + Maosong Sun Tsinghua + OpenBMB + ModelScope authors + Industry-leading ToolLLM GitHub OpenBMB/ToolBench open source + Industry-leading 16000+ real-world API LLM Tsinghua 2023 position確立。