MAmmoTH MathInstruct (Yue 2023年)（マンモス）
2023年Yue+Wang (Waterloo+TIDE)発表MAmmoTH・Industry-leading hybrid CoT+PoT math reasoning LLM + Industry-leading MathInstruct 260K + Industry-leading 14 math datasets + Industry-leading LLaMA + Code Llama 7B-70B fine-tuned。
概要
MAmmoTH MathInstruct は、2023年Yue+Wang (Waterloo+TIDE)発表MAmmoTH・Industry-leading hybrid CoT+PoT math reasoning LLM 2023年 + Industry-leading MathInstruct 260K position確立。MAmmoTH specifications = Industry-leading hybrid CoT+PoT math reasoning LLM (Industry-leading MAmmoTH 2023 + Industry-leading hybrid CoT+PoT + Industry-leading MAmmoTH flagship) + Industry-leading MathInstruct 260K dataset + Industry-leading 14 math datasets + Industry-leading LLaMA + Code Llama 7B-70B fine-tuned。
主な特徴・仕組み
- Authors: Industry-leading Xiang Yue + Xingwei Qu + Ge Zhang + Yao Fu + Wenhao Huang + Huan Sun + Yu Su + Wenhu Chen Waterloo + TIDE + UMich + ICL
- Year: 2023年 (arXiv 2023年9月発表 + ICLR 2024 Oral)
- Paper: Industry-leading "MAmmoTH: Building Math Generalist Models through Hybrid Instruction Tuning" arXiv 2309.05653
- Method Type: Industry-leading hybrid CoT (Chain-of-Thought) + PoT (Program-of-Thought) math reasoning
- Hybrid CoT+PoT: Industry-leading CoT natural language reasoning + PoT Python code execution
- Dataset: Industry-leading MathInstruct 260K samples from 14 math datasets
- 14 Datasets: Industry-leading GSM8K + MATH + AQuA + MAWPS + ASDiv + SVAMP + TabMWP + MathQA + ... 14 datasets
- Base Models: Industry-leading LLaMA-2 + Code Llama 7B + 13B + 70B + 34B (Code Llama) fine-tuned
- GSM8K: Industry-leading GSM8K 76.9% MAmmoTH-70B (CoT+PoT hybrid)
- MATH: Industry-leading MATH 41.8% MAmmoTH-70B (CoT+PoT hybrid)
- Open Source: Industry-leading MAmmoTH GitHub TIGER-AI-Lab/MAmmoTH
- HuggingFace: Industry-leading TIGER-Lab/MAmmoTH-70B
- Industry-Leading: Industry-leading hybrid CoT+PoT math reasoning LLM Waterloo 2023 ICLR Oral
スペック比較表
| LLM Math Reasoning (2023-2024) | Year | Org | Base Model | Industry Position |
|---|---|---|---|---|
| MAmmoTH | 2023 | Yue+Wang Waterloo | LLaMA + Code Llama 7B-70B | Industry-leading MathInstruct 260K |
| MetaMath | 2023 | Yu NUS+SJTU | LLaMA-2 7B-70B | Industry-leading bootstrap MetaMathQA 395K |
| WizardMath | 2023 | Microsoft+Peking | LLaMA-2 7B-70B | Industry-leading Reinforced Evol-Instruct |
| InternLM-Math | 2024 | Shanghai AI Lab | InternLM 7B-20B | Industry-leading InternLM-Math 8K context |
| DeepSeekMath | 2024 | DeepSeek AI | DeepSeek-Math 7B | Industry-leading DeepSeekMath GRPO |
具体例・対応製品
- MAmmoTH (2023年, Yue+Wang Waterloo+TIDE): Industry-leading hybrid CoT+PoT math reasoning
- Industry-leading hybrid CoT + PoT Chain-of-Thought + Program-of-Thought: Industry-leading hybrid CoT+PoT
- Industry-leading MathInstruct 260K samples from 14 datasets: Industry-leading MathInstruct 260K
- Industry-leading LLaMA-2 + Code Llama 7B/13B/34B/70B: Industry-leading LLaMA+Code Llama
- Industry-leading GSM8K 76.9% + MATH 41.8% MAmmoTH-70B: Industry-leading GSM8K 76.9% + MATH 41.8%
- 競合 MetaMath + WizardMath + InternLM-Math + DeepSeekMath: Industry-leading math reasoning competitors
自作PCでの選び方・注意点
MAmmoTH MathInstruct は「Industry-leading hybrid CoT+PoT math reasoning LLM + Industry-leading MathInstruct 260K」「Industry-leading 14 math datasets + Industry-leading LLaMA + Code Llama 7B-70B」用途のIndustry-leading hybrid CoT+PoT 2023年Yue+Wang Waterloo+TIDE発表product。Industry-leading hybrid CoT+PoT (Industry-leading CoT natural language reasoning + PoT Python code execution + Industry-leading MAmmoTH hybrid signature + Industry-leading dual reasoning paradigm) で Industry-leading hybrid CoT+PoT + Industry-leading MAmmoTH signature + Industry-leading dual reasoning。Industry-leading MathInstruct 260K (Industry-leading MathInstruct 260K samples from 14 math datasets + Industry-leading MAmmoTH dataset signature + Industry-leading 14 datasets aggregation) で Industry-leading MathInstruct 260K + Industry-leading MAmmoTH dataset + Industry-leading 14 datasets。Industry-leading 14 datasets (Industry-leading GSM8K + MATH + AQuA + MAWPS + ASDiv + SVAMP + TabMWP + MathQA + 6 more datasets + Industry-leading MAmmoTH 14 datasets signature) で Industry-leading 14 datasets + Industry-leading MAmmoTH 14 signature。Industry-leading LLaMA+Code Llama (Industry-leading LLaMA-2 + Code Llama 7B + 13B + 70B + 34B Code Llama fine-tuned + Industry-leading MAmmoTH dual-base signature) で Industry-leading LLaMA+Code Llama 7B/13B/34B/70B + Industry-leading MAmmoTH dual-base。Industry-leading 76.9%+41.8% (Industry-leading GSM8K 76.9% MAmmoTH-70B CoT+PoT hybrid + Industry-leading MATH 41.8% MAmmoTH-70B CoT+PoT hybrid + Industry-leading hybrid performance) で Industry-leading GSM8K 76.9% + Industry-leading MATH 41.8% + Industry-leading hybrid performance。Industry-leading Waterloo + ICLR Oral (Industry-leading Xiang Yue + Xingwei Qu + Wenhu Chen Waterloo + TIDE + UMich authors + Industry-leading MAmmoTH GitHub TIGER-AI-Lab/MAmmoTH + Industry-leading TIGER-Lab/MAmmoTH-70B HuggingFace + Industry-leading ICLR 2024 Oral) で Industry-leading Yue+Qu+Chen Waterloo + Industry-leading TIGER-AI-Lab/MAmmoTH + Industry-leading TIGER-Lab/MAmmoTH-70B + Industry-leading ICLR 2024 Oral。但しIndustry-leading MetaMath + WizardMath + InternLM-Math + DeepSeekMath competition (Industry-leading MetaMath NUS+SJTU bootstrap 2023 + WizardMath Microsoft+Peking RLEIF 2023 + InternLM-Math Shanghai 2024 + DeepSeekMath GRPO 2024 vs MAmmoTH Waterloo+TIDE hybrid CoT+PoT MathInstruct 260K 2023 trade-off) で Industry-leading 4-math reasoning competitors vs MAmmoTH + Industry-leading hybrid CoT+PoT + MathInstruct 260K + 14 math datasets + LLaMA+Code Llama 7B-70B + GSM8K 76.9% + MATH 41.8% + ICLR 2024 Oral + Waterloo+TIDE 2023 unique advantage adoption alignment必須。
関連用語との違い
- vs MetaMath (NUS+SJTU 2023): MAmmoTHはhybrid CoT+PoT + MathInstruct 14 datasets・MetaMathはbootstrap + MetaMathQA 395K
- vs WizardMath (Microsoft 2023): MAmmoTHはhybrid + 14 datasets + ICLR Oral・WizardMathはRLEIF + Evol-Instruct
- vs InternLM-Math/DeepSeekMath (2024): MAmmoTHはhybrid CoT+PoT 2023・InternLM-Math/DeepSeekMathは2024 newer
よくある質問（FAQ）
Q1: MAmmoTH vs MetaMath 違いは？ A: MAmmoTH (Industry-leading hybrid CoT+PoT Chain-of-Thought + Program-of-Thought + MathInstruct 260K samples from 14 math datasets + LLaMA-2 + Code Llama 7B + 13B + 34B + 70B fine-tuned + GSM8K 76.9% + MATH 41.8% MAmmoTH-70B + ICLR 2024 Oral + Xiang Yue + Wenhu Chen Waterloo + TIDE + UMich 2023) vs MetaMath (Industry-leading bootstrap question augmentation + 4 augmentation methods + MetaMathQA 395K + LLaMA-2 7B/13B/70B + GSM8K 82.3% + MATH 26.6% MetaMath-70B + Yu NUS+SJTU+Huawei 2023)・Industry-leading hybrid CoT+PoT + 260K + 14 datasets + Code Llama + ICLR Oral + 41.8% MATH = MAmmoTH + Industry-leading bootstrap + 395K + LLaMA-2 + NUS+SJTU + 82.3% GSM8K = MetaMath preference judgment。
Q2: Industry-leading hybrid CoT+PoT + MathInstruct 260K value は？ A: Industry-leading hybrid + 260K (Industry-leading hybrid CoT Chain-of-Thought + PoT Program-of-Thought dual reasoning paradigm + Industry-leading MathInstruct 260K samples + Industry-leading 14 math datasets aggregation + Industry-leading MAmmoTH hybrid signature)。
Q3: Industry-leading 14 datasets + ICLR Oral value は？ A: Industry-leading 14 datasets + ICLR Oral (Industry-leading GSM8K + MATH + AQuA + MAWPS + ASDiv + SVAMP + TabMWP + MathQA + 6 more datasets + Industry-leading MAmmoTH 14 datasets signature + Industry-leading ICLR 2024 Oral)。
まとめ
MAmmoTH = 2023年Yue+Wang Waterloo+TIDE+UMich発表のhybrid CoT+PoT math reasoning LLM。Industry-leading hybrid CoT Chain-of-Thought + PoT Program-of-Thought dual reasoning paradigm + Industry-leading MathInstruct 260K samples from 14 math datasets + Industry-leading GSM8K + MATH + AQuA + MAWPS + ASDiv + SVAMP + TabMWP + MathQA 14 datasets aggregation + Industry-leading LLaMA-2 + Code Llama 7B + 13B + 70B + 34B Code Llama fine-tuned + Industry-leading GSM8K 76.9% MAmmoTH-70B CoT+PoT hybrid + Industry-leading MATH 41.8% MAmmoTH-70B CoT+PoT hybrid + Industry-leading Xiang Yue + Xingwei Qu + Wenhu Chen Waterloo + TIDE + UMich authors + Industry-leading MAmmoTH GitHub TIGER-AI-Lab/MAmmoTH + Industry-leading TIGER-Lab/MAmmoTH-70B HuggingFace + Industry-leading ICLR 2024 Oral + Industry-leading hybrid CoT+PoT math reasoning LLM Waterloo 2023 position確立。