Phi-3 textbooks (Microsoft 2024年)（ファイスリーテキストブックス）
2024年Microsoft発表Phi-3 textbooks・Industry-leading textbook quality synthetic data + Industry-leading "Textbooks Are All You Need" methodology + Industry-leading Phi-3-mini 3.8B SLM + Industry-leading high-quality data > scale paradigm。
概要
Phi-3 textbooks は、2024年Microsoft発表Phi-3 textbooks・Industry-leading textbook quality synthetic data 2024年 + Industry-leading high-quality data > scale paradigm position確立。Phi-3 textbooks specifications = Industry-leading textbook quality synthetic data (Industry-leading Phi-3 textbooks Microsoft 2024 + Industry-leading textbook quality + Industry-leading Phi-3 flagship) + Industry-leading "Textbooks Are All You Need" methodology + Industry-leading Phi-3-mini 3.8B SLM + Industry-leading high-quality data > scale paradigm。
主な特徴・仕組み
- Org: Industry-leading Microsoft Research (Sébastien Bubeck + Ronen Eldan + Yuanzhi Li + Allie Del Giorno + Suriya Gunasekar lab)
- Year: 2024年 (Phi-3 release 2024年4月)
- Paper Series: Industry-leading "Textbooks Are All You Need" Phi-1 2023年6月 + Phi-1.5 2023年9月 + Phi-2 2023年12月 + Phi-3 2024年4月 + Phi-3.5 2024年8月
- Phi-3 Paper: Industry-leading "Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone" arXiv 2404.14219
- Method Type: Industry-leading textbook quality synthetic data SLM
- Textbooks Methodology: Industry-leading "Textbooks Are All You Need" methodology
- Phi-3-mini: Industry-leading Phi-3-mini 3.8B parameters SLM
- Phi-3 Models: Industry-leading Phi-3-mini 3.8B + Phi-3-small 7B + Phi-3-medium 14B
- Quality > Scale: Industry-leading high-quality data > scale paradigm signature
- Edu Score Filter: Industry-leading Edu Score Filter data filtering
- Mobile Local: Industry-leading mobile local SLM deployment
- Industry-Leading: Industry-leading textbook quality synthetic data SLM Microsoft 2024
スペック比較表
| LLM Synthetic Data (2022-2024) | Year | Org | Method | Industry Position |
|---|---|---|---|---|
| Phi-3 textbooks | 2024 | Microsoft | Textbook quality data | Industry-leading Phi-3 textbooks |
| Self-Instruct | 2022 | Wang et al. UW+AI2 | Self-generated 52K | Industry-leading synthetic foundation |
| Alpaca | 2023 | Taori et al. Stanford | 52K Alpaca dataset | Industry-leading Stanford Alpaca |
| Orca | 2023 | Mukherjee Microsoft | Explanation tuning | Industry-leading Microsoft Orca |
| WizardLM Evol-Instruct | 2023 | Xu et al. Microsoft+Peking | Evolved instructions | Industry-leading Evol-Instruct |
具体例・対応製品
- Phi-3 textbooks (2024年, Microsoft Research): Industry-leading textbook quality + Phi-3-mini 3.8B SLM
- Industry-leading "Textbooks Are All You Need" methodology: Industry-leading Textbooks Are All You Need
- Industry-leading Phi-3-mini 3.8B parameters SLM: Industry-leading Phi-3-mini 3.8B
- Industry-leading Phi-3-mini + Phi-3-small 7B + Phi-3-medium 14B: Industry-leading Phi-3 model family
- Industry-leading high-quality data > scale paradigm: Industry-leading quality > scale
- 競合 Self-Instruct + Alpaca + Orca + WizardLM Evol-Instruct: Industry-leading synthetic data competitors
選び方・注意点
Phi-3 textbooks は「Industry-leading textbook quality synthetic data + 'Textbooks Are All You Need'」「Industry-leading Phi-3-mini 3.8B SLM + quality > scale」用途のIndustry-leading 2024年 Microsoft Research product。Industry-leading textbook quality (Industry-leading textbook quality synthetic data + Industry-leading Phi-3 textbook quality signature + Industry-leading high-quality data curation) で Industry-leading textbook quality + Industry-leading Phi-3 signature + Industry-leading high-quality curation。Industry-leading "Textbooks Are All You Need" (Industry-leading "Textbooks Are All You Need" methodology + Industry-leading Phi-series methodology signature + Industry-leading Phi-1 2023年6月 + Phi-1.5 2023年9月 + Phi-2 2023年12月 + Phi-3 2024年4月 + Phi-3.5 2024年8月 progression) で Industry-leading "Textbooks Are All You Need" + Industry-leading Phi-series signature + Industry-leading Phi-1/1.5/2/3/3.5 progression。Industry-leading Phi-3-mini 3.8B (Industry-leading Phi-3-mini 3.8B parameters SLM + Industry-leading Phi-3-mini signature + Industry-leading small language model SLM efficiency) で Industry-leading Phi-3-mini 3.8B + Industry-leading Phi-3 signature + Industry-leading SLM efficiency。Industry-leading quality > scale (Industry-leading high-quality data > scale paradigm signature + Industry-leading Phi-3 paradigm signature + Industry-leading data-centric AI shift) で Industry-leading quality > scale + Industry-leading Phi-3 paradigm + Industry-leading data-centric AI。Industry-leading Phi-3 family (Industry-leading Phi-3-mini 3.8B + Phi-3-small 7B + Phi-3-medium 14B + Industry-leading Phi-3 model family + Industry-leading Sébastien Bubeck + Ronen Eldan + Yuanzhi Li + Suriya Gunasekar Microsoft Research lab + Industry-leading mobile local SLM deployment + Industry-leading Edu Score Filter data filtering) で Industry-leading Phi-3 family + Industry-leading Bubeck+Eldan+Li+Gunasekar Microsoft + Industry-leading mobile SLM + Industry-leading Edu Score Filter。但しIndustry-leading Self-Instruct + Alpaca + Orca + WizardLM Evol-Instruct competition で Industry-leading 4-synthetic data competitors vs Phi-3 textbooks + Industry-leading textbook quality synthetic data + "Textbooks Are All You Need" methodology + Phi-3-mini 3.8B + Phi-3-small 7B + Phi-3-medium 14B + high-quality data > scale paradigm + Edu Score Filter + mobile local SLM deployment + Sébastien Bubeck + Ronen Eldan + Yuanzhi Li + Suriya Gunasekar Microsoft Research 2024 unique advantage adoption alignment必須。
関連用語との違い
- vs Self-Instruct/Alpaca/WizardLM (Instruction tuning 2022-2023): Phi-3 textbooksはpre-training synthetic data + quality > scale・他はinstruction-tuning synthetic data
- vs Orca (Microsoft 2023): Phi-3 textbooksはtextbook quality pre-training Microsoft 2024・OrcaはGPT-4 explanation tuning Microsoft 2023
- vs Phi-1/Phi-2 (Microsoft 2023): Phi-3 textbooksは2024 Phi-3-mini 3.8B + Phi-3-medium 14B latest・Phi-1/Phi-2は2023 predecessors
よくある質問（FAQ）
Q1: Phi-3 textbooks vs Orca 違いは？ A: Phi-3 textbooks (Industry-leading textbook quality synthetic data + "Textbooks Are All You Need" methodology + Phi-3-mini 3.8B + Phi-3-small 7B + Phi-3-medium 14B + high-quality data > scale paradigm + Edu Score Filter + mobile local SLM + Sébastien Bubeck + Ronen Eldan + Yuanzhi Li Microsoft Research 2024) vs Orca (Industry-leading explanation tuning + GPT-4 teacher model + step-by-step reasoning trace + Orca 13B + Orca 2 7B + 13B + Mukherjee + Mitra + Palangi + Awadallah Microsoft Research 2023)・Industry-leading textbook quality + Phi-series + Phi-3-mini 3.8B + quality > scale + Bubeck + 2024 = Phi-3 textbooks + Industry-leading explanation tuning + GPT-4 teacher + Orca 13B + Mukherjee + 2023 = Orca preference judgment (pre-training quality vs instruction explanation tuning differ)。
Q2: Industry-leading "Textbooks Are All You Need" + Phi-3-mini 3.8B value は？ A: Industry-leading Textbooks Are All You Need + Phi-3-mini (Industry-leading "Textbooks Are All You Need" methodology + Industry-leading Phi-3-mini 3.8B parameters SLM + Industry-leading Phi-3 dual signature)。
Q3: Industry-leading quality > scale paradigm + Phi-3 family value は？ A: Industry-leading quality > scale + Phi-3 family (Industry-leading high-quality data > scale paradigm signature + Industry-leading Phi-3-mini 3.8B + Phi-3-small 7B + Phi-3-medium 14B + Industry-leading Bubeck+Eldan+Li+Gunasekar Microsoft Research lab)。
まとめ
Phi-3 textbooks = 2024年Microsoft Research発表のtextbook quality synthetic data SLM methodology。Industry-leading textbook quality synthetic data + Industry-leading "Textbooks Are All You Need" methodology + Industry-leading Phi-3-mini 3.8B + Phi-3-small 7B + Phi-3-medium 14B model family + Industry-leading high-quality data > scale paradigm + Industry-leading Edu Score Filter data filtering + Industry-leading mobile local SLM deployment + Industry-leading Phi-1 2023年6月 + Phi-1.5 2023年9月 + Phi-2 2023年12月 + Phi-3 2024年4月 + Phi-3.5 2024年8月 progression + Industry-leading Sébastien Bubeck + Ronen Eldan + Yuanzhi Li + Allie Del Giorno + Suriya Gunasekar Microsoft Research authors + Industry-leading "Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone" arXiv 2404.14219 + Industry-leading textbook quality synthetic data SLM Microsoft 2024 position確立。