InternLM-Math (Shanghai AI Lab 2024年)(インターンエルエムマス)
2024年Shanghai AI Lab発表InternLM-Math・Industry-leading bilingual math reasoning LLM + Industry-leading InternLM 7B/20B base + Industry-leading 8K context + Industry-leading Chinese+English math support。
概要
InternLM-Math は、2024年Shanghai AI Lab発表InternLM-Math・Industry-leading bilingual math reasoning LLM 2024年 + Industry-leading Chinese+English math support position確立。InternLM-Math specifications = Industry-leading bilingual math reasoning LLM (Industry-leading InternLM-Math 2024 + Industry-leading bilingual math reasoning + Industry-leading InternLM-Math flagship) + Industry-leading InternLM 7B/20B base + Industry-leading 8K context + Industry-leading Chinese+English math support。
主な特徴・仕組み
- Org: Industry-leading Shanghai AI Lab (上海人工智能实验室)
- Year: 2024年 (arXiv 2024年2月発表)
- Paper: Industry-leading "InternLM-Math: Open Math Large Language Models Toward Verifiable Reasoning" arXiv 2402.06332
- Sizes: Industry-leading InternLM-Math 7B + 20B variants
- Base Model: Industry-leading InternLM 7B/20B base fine-tuned for math
- Context Length: Industry-leading 8K tokens context
- Bilingual: Industry-leading Chinese + English math support signature
- Verifiable Reasoning: Industry-leading Lean 4 + Coq formal verification support
- GSM8K: Industry-leading GSM8K 84.6% InternLM-Math-20B
- MATH: Industry-leading MATH 37.7% InternLM-Math-20B
- Open Source: Industry-leading InternLM-Math GitHub InternLM/InternLM-Math
- HuggingFace: Industry-leading internlm/internlm2-math-20b
- Industry-Leading: Industry-leading bilingual math reasoning LLM Shanghai AI Lab 2024
スペック比較表
| LLM Math Reasoning (2023-2024) | Year | Org | Base Model | Industry Position |
|---|---|---|---|---|
| InternLM-Math | 2024 | Shanghai AI Lab | InternLM 7B-20B | Industry-leading InternLM-Math bilingual |
| DeepSeekMath | 2024 | DeepSeek AI | DeepSeek-Math 7B | Industry-leading DeepSeekMath GRPO |
| 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 |
| MAmmoTH | 2023 | Yue+Wang Waterloo | LLaMA + Code Llama 7B-70B | Industry-leading MathInstruct 260K |
具体例・対応製品
- InternLM-Math (2024年, Shanghai AI Lab): Industry-leading bilingual math reasoning LLM
- Industry-leading InternLM 7B + 20B base: Industry-leading InternLM 7B/20B
- Industry-leading Chinese + English math bilingual signature: Industry-leading Chinese+English bilingual
- Industry-leading Lean 4 + Coq formal verification: Industry-leading Lean 4+Coq formal verification
- Industry-leading GSM8K 84.6% + MATH 37.7% InternLM-Math-20B: Industry-leading GSM8K 84.6% + MATH 37.7%
- 競合 MetaMath + WizardMath + MAmmoTH + DeepSeekMath: Industry-leading math reasoning competitors
自作PCでの選び方・注意点
InternLM-Math は「Industry-leading bilingual math reasoning LLM + Industry-leading InternLM 7B/20B base」「Industry-leading 8K context + Industry-leading Chinese+English math support」用途のIndustry-leading bilingual math reasoning 2024年Shanghai AI Lab発表product。Industry-leading InternLM base (Industry-leading InternLM 7B + 20B base fine-tuned for math + Industry-leading InternLM-Math InternLM-base signature) で Industry-leading InternLM 7B/20B + Industry-leading InternLM-Math signature。Industry-leading bilingual (Industry-leading Chinese + English math support signature + Industry-leading InternLM-Math bilingual signature + Industry-leading dual-language math) で Industry-leading bilingual + Industry-leading InternLM-Math bilingual + Industry-leading dual-language math。Industry-leading verifiable reasoning (Industry-leading Lean 4 + Coq formal verification support + Industry-leading verifiable reasoning signature + Industry-leading formal math verification) で Industry-leading Lean 4+Coq + Industry-leading verifiable signature + Industry-leading formal verification。Industry-leading 8K context (Industry-leading 8K tokens context + Industry-leading InternLM-Math 8K context signature + Industry-leading long context math) で Industry-leading 8K + Industry-leading InternLM-Math 8K + Industry-leading long context math。Industry-leading GSM8K 84.6% + MATH 37.7% (Industry-leading GSM8K 84.6% InternLM-Math-20B + Industry-leading MATH 37.7% InternLM-Math-20B + Industry-leading InternLM-Math-20B performance) で Industry-leading GSM8K 84.6% + Industry-leading MATH 37.7% + Industry-leading 20B performance。Industry-leading Shanghai AI Lab (Industry-leading Shanghai AI Lab 上海人工智能实验室 + Industry-leading InternLM-Math GitHub InternLM/InternLM-Math + Industry-leading internlm/internlm2-math-20b HuggingFace) で Industry-leading Shanghai AI Lab + Industry-leading InternLM/InternLM-Math + Industry-leading internlm/internlm2-math-20b。但しIndustry-leading MetaMath + WizardMath + MAmmoTH + DeepSeekMath competition (Industry-leading MetaMath NUS+SJTU 2023 + WizardMath Microsoft 2023 + MAmmoTH Waterloo 2023 + DeepSeekMath GRPO 2024 vs InternLM-Math Shanghai AI Lab bilingual 8K context Lean 4 verification 2024 trade-off) で Industry-leading 4-math reasoning competitors vs InternLM-Math + Industry-leading bilingual Chinese+English + InternLM 7B/20B + 8K context + Lean 4 + Coq formal verification + GSM8K 84.6% + MATH 37.7% + Shanghai AI Lab 2024 unique advantage adoption alignment必須。
関連用語との違い
- vs MetaMath/WizardMath/MAmmoTH (2023): InternLM-MathはInternLM base + bilingual + Lean 4 verification + 2024・他はLLaMA-2 base + 2023
- vs DeepSeekMath (DeepSeek 2024): InternLM-MathはInternLM 7B/20B + bilingual + Shanghai・DeepSeekMathはDeepSeek-Math 7B + GRPO + DeepSeek
- vs General LLM (no math fine-tuning): InternLM-Mathはmath-specialized + bilingual + Lean 4・GeneralはGeneral-purpose
よくある質問(FAQ)
Q1: InternLM-Math vs DeepSeekMath 違いは? A: InternLM-Math (Industry-leading InternLM 7B + 20B base fine-tuned + 8K tokens context + Chinese + English math bilingual support + Lean 4 + Coq formal verification + GSM8K 84.6% + MATH 37.7% InternLM-Math-20B + Shanghai AI Lab 2024) vs DeepSeekMath (Industry-leading DeepSeek-Math 7B + GRPO Group Relative Policy Optimization + 120B Common Crawl math + GSM8K 64.2% + MATH 51.7% + DeepSeek AI 2024)・Industry-leading InternLM base + bilingual + Lean 4 verification + 7B/20B + Shanghai = InternLM-Math + Industry-leading DeepSeek-Math + GRPO + 120B Common Crawl + 51.7% MATH + DeepSeek = DeepSeekMath preference judgment。
Q2: Industry-leading bilingual Chinese+English + InternLM base value は? A: Industry-leading bilingual + InternLM (Industry-leading Chinese + English math support signature + Industry-leading InternLM-Math bilingual signature + Industry-leading InternLM 7B + 20B base fine-tuned for math + Industry-leading dual-language math)。
Q3: Industry-leading Lean 4 + Coq verifiable reasoning value は? A: Industry-leading Lean 4 + Coq (Industry-leading Lean 4 + Coq formal verification support + Industry-leading verifiable reasoning signature + Industry-leading formal math verification + Industry-leading InternLM-Math verifiable signature)。
まとめ
InternLM-Math = 2024年Shanghai AI Lab発表のbilingual math reasoning LLM。Industry-leading InternLM-Math 7B + 20B variants + Industry-leading InternLM 7B/20B base fine-tuned for math + Industry-leading 8K tokens context + Industry-leading Chinese + English math support signature + Industry-leading Lean 4 + Coq formal verification support + Industry-leading verifiable reasoning signature + Industry-leading GSM8K 84.6% InternLM-Math-20B + Industry-leading MATH 37.7% InternLM-Math-20B + Industry-leading Shanghai AI Lab 上海人工智能实验室 + Industry-leading InternLM-Math GitHub InternLM/InternLM-Math + Industry-leading internlm/internlm2-math-20b HuggingFace + Industry-leading bilingual math reasoning LLM Shanghai AI Lab 2024 position確立。