WizardMath Reinforced Evol-Instruct (Microsoft 2023年)（ウィザードマス）
2023年Microsoft+Peking発表WizardMath・Industry-leading Reinforced Evol-Instruct math reasoning LLM + Industry-leading RLEIF + Industry-leading LLaMA-2 7B-70B fine-tuned + Industry-leading GSM8K 81.6% MATH 22.7%。
概要
WizardMath Reinforced Evol-Instruct は、2023年Microsoft+Peking発表WizardMath・Industry-leading Reinforced Evol-Instruct math reasoning LLM 2023年 + Industry-leading RLEIF position確立。WizardMath specifications = Industry-leading Reinforced Evol-Instruct math reasoning LLM (Industry-leading WizardMath 2023 + Industry-leading Reinforced Evol-Instruct + Industry-leading WizardMath flagship) + Industry-leading RLEIF (Reinforcement Learning from Evol-Instruct Feedback) + Industry-leading LLaMA-2 7B-70B fine-tuned + Industry-leading GSM8K 81.6% MATH 22.7%。
主な特徴・仕組み
- Authors: Industry-leading Haipeng Luo + Qingfeng Sun + Can Xu + Pu Zhao + Jianguang Lou + Chongyang Tao + Xiubo Geng + Qingwei Lin + Shifeng Chen + Dongmei Zhang Microsoft + Peking
- Year: 2023年 (arXiv 2023年8月発表)
- Paper: Industry-leading "WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct" arXiv 2308.09583
- Method Type: Industry-leading Reinforced Evol-Instruct math reasoning LLM
- RLEIF: Industry-leading Reinforcement Learning from Evol-Instruct Feedback signature
- Evol-Instruct: Industry-leading Evol-Instruct evolved instruction methodology
- Base Model: Industry-leading LLaMA-2 7B + 13B + 70B fine-tuned
- GSM8K: Industry-leading GSM8K 81.6% pass@1 WizardMath-70B
- MATH: Industry-leading MATH 22.7% pass@1 WizardMath-70B
- WizardLM Series: Industry-leading WizardLM family (WizardMath + WizardCoder + WizardLM)
- Open Source: Industry-leading WizardMath GitHub nlpxucan/WizardLM
- HuggingFace: Industry-leading WizardLMTeam/WizardMath-70B-V1.0
- Industry-Leading: Industry-leading Reinforced Evol-Instruct WizardMath Microsoft+Peking 2023
スペック比較表
| LLM Math Reasoning (2023-2024) | Year | Org | Base Model | Industry Position |
|---|---|---|---|---|
| WizardMath | 2023 | Microsoft+Peking | LLaMA-2 7B-70B | Industry-leading Reinforced Evol-Instruct |
| MetaMath | 2023 | Yu NUS+SJTU | LLaMA-2 7B-70B | Industry-leading bootstrap MetaMathQA 395K |
| MAmmoTH | 2023 | Yue+Wang Waterloo | LLaMA + Code Llama 7B-70B | Industry-leading MathInstruct 260K |
| 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
- WizardMath (2023年, Microsoft+Peking): Industry-leading Reinforced Evol-Instruct math reasoning
- Industry-leading RLEIF Reinforcement Learning from Evol-Instruct Feedback: Industry-leading RLEIF signature
- Industry-leading Evol-Instruct evolved instruction methodology: Industry-leading Evol-Instruct
- Industry-leading LLaMA-2 7B/13B/70B fine-tuned: Industry-leading LLaMA-2 base
- Industry-leading WizardLM family WizardMath+WizardCoder+WizardLM: Industry-leading WizardLM family
- 競合 MetaMath + MAmmoTH + InternLM-Math + DeepSeekMath: Industry-leading math reasoning competitors
自作PCでの選び方・注意点
WizardMath Reinforced Evol-Instruct は「Industry-leading Reinforced Evol-Instruct math reasoning LLM + Industry-leading RLEIF」「Industry-leading LLaMA-2 7B-70B fine-tuned + Industry-leading GSM8K 81.6% MATH 22.7%」用途のIndustry-leading Reinforced Evol-Instruct 2023年Microsoft+Peking発表product。Industry-leading RLEIF (Industry-leading Reinforcement Learning from Evol-Instruct Feedback signature + Industry-leading WizardMath RLEIF heritage + Industry-leading RL fine-tuning) で Industry-leading RLEIF + Industry-leading WizardMath RLEIF + Industry-leading RL fine-tuning。Industry-leading Evol-Instruct (Industry-leading Evol-Instruct evolved instruction methodology + Industry-leading WizardMath Evol-Instruct signature + Industry-leading instruction evolution) で Industry-leading Evol-Instruct + Industry-leading WizardMath signature + Industry-leading instruction evolution。Industry-leading LLaMA-2 (Industry-leading LLaMA-2 7B + 13B + 70B fine-tuned + Industry-leading WizardMath LLaMA-2 signature) で Industry-leading LLaMA-2 7B/13B/70B + Industry-leading WizardMath LLaMA-2 signature。Industry-leading GSM8K 81.6% + MATH 22.7% (Industry-leading GSM8K 81.6% pass@1 WizardMath-70B + Industry-leading MATH 22.7% pass@1 WizardMath-70B + Industry-leading WizardMath-70B performance) で Industry-leading GSM8K 81.6% + Industry-leading MATH 22.7% + Industry-leading WizardMath-70B performance。Industry-leading WizardLM family (Industry-leading WizardLM family WizardMath + WizardCoder + WizardLM + Industry-leading WizardLM ecosystem + Industry-leading Microsoft+Peking WizardLM heritage) で Industry-leading WizardLM family + Industry-leading WizardLM ecosystem + Industry-leading Microsoft+Peking heritage。但しIndustry-leading MetaMath + MAmmoTH + InternLM-Math + DeepSeekMath competition (Industry-leading MetaMath NUS+SJTU bootstrap 2023 + MAmmoTH Waterloo MathInstruct 2023 + InternLM-Math Shanghai 2024 + DeepSeekMath GRPO 2024 vs WizardMath Microsoft+Peking RLEIF Evol-Instruct 2023 trade-off) で Industry-leading 4-math reasoning competitors vs WizardMath + Industry-leading Reinforced Evol-Instruct RLEIF + Evol-Instruct methodology + LLaMA-2 7B-70B + GSM8K 81.6% + MATH 22.7% + WizardLM family + Microsoft+Peking 2023 unique advantage adoption alignment必須。
関連用語との違い
- vs MetaMath (NUS+SJTU 2023): WizardMathはRLEIF + Evol-Instruct + Microsoft・MetaMathはbootstrap 4 augmentation + MetaMathQA 395K + NUS
- vs MAmmoTH (Waterloo 2023): WizardMathはRLEIF + WizardLM family・MAmmoTHはMathInstruct 260K + 14 datasets
- vs DeepSeekMath (DeepSeek 2024): WizardMathはLLaMA-2 + Microsoft・DeepSeekMathはDeepSeek-Math + GRPO
よくある質問（FAQ）
Q1: WizardMath vs MetaMath 違いは？ A: WizardMath (Industry-leading Reinforced Evol-Instruct RLEIF + Evol-Instruct evolved instruction + LLaMA-2 7B/13B/70B + GSM8K 81.6% pass@1 + MATH 22.7% WizardMath-70B + WizardLM family WizardMath+WizardCoder+WizardLM + Luo + Sun + Xu Microsoft + Peking 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 RLEIF + Evol-Instruct + WizardLM family + Microsoft+Peking + 81.6% = WizardMath + Industry-leading bootstrap + 4 augmentation + MetaMathQA + NUS+SJTU + 82.3% = MetaMath preference judgment。
Q2: Industry-leading RLEIF + Evol-Instruct value は？ A: Industry-leading RLEIF + Evol-Instruct (Industry-leading Reinforcement Learning from Evol-Instruct Feedback signature + Industry-leading Evol-Instruct evolved instruction methodology + Industry-leading WizardMath RLEIF heritage + Industry-leading RL fine-tuning + instruction evolution)。
Q3: Industry-leading WizardLM family + Microsoft+Peking value は？ A: Industry-leading WizardLM family + Microsoft+Peking (Industry-leading WizardLM family WizardMath + WizardCoder + WizardLM + Industry-leading WizardLM ecosystem + Industry-leading Microsoft+Peking WizardLM heritage + Industry-leading Haipeng Luo + Qingfeng Sun Microsoft + Peking authors)。
まとめ
WizardMath = 2023年Microsoft+Peking発表のReinforced Evol-Instruct math reasoning LLM。Industry-leading Reinforcement Learning from Evol-Instruct Feedback RLEIF signature + Industry-leading Evol-Instruct evolved instruction methodology + Industry-leading LLaMA-2 7B + 13B + 70B fine-tuned + Industry-leading GSM8K 81.6% pass@1 WizardMath-70B + Industry-leading MATH 22.7% pass@1 WizardMath-70B + Industry-leading WizardLM family WizardMath + WizardCoder + WizardLM + Industry-leading Haipeng Luo + Qingfeng Sun + Can Xu + Dongmei Zhang Microsoft + Peking authors + Industry-leading WizardMath GitHub nlpxucan/WizardLM + Industry-leading WizardLMTeam/WizardMath-70B-V1.0 HuggingFace + Industry-leading Reinforced Evol-Instruct WizardMath Microsoft+Peking 2023 position確立。