BGE Large Embedding (BAAI 2023年)（ビージーイーラージ）
2023年BAAI (Beijing Academy of AI)発表BGE Large・Industry-leading open-source embedding LLM + Industry-leading 1024-dim BERT-large + Industry-leading MTEB benchmark top + Industry-leading Chinese+English bilingual embedding。
概要
BGE Large Embedding は、2023年BAAI (Beijing Academy of AI)発表BGE Large・Industry-leading open-source embedding LLM 2023年 + Industry-leading MTEB benchmark top position確立。BGE Large specifications = Industry-leading open-source embedding LLM (Industry-leading BGE Large 2023 + Industry-leading open-source embedding + Industry-leading BGE flagship) + Industry-leading 1024-dim BERT-large (Industry-leading 1024-dim BERT-large embedding) + Industry-leading MTEB benchmark top (Industry-leading MTEB benchmark top 2023) + Industry-leading Chinese+English bilingual embedding (Industry-leading Chinese+English bilingual embedding)。
主な特徴・仕組み
- Org: Industry-leading BAAI (Beijing Academy of Artificial Intelligence)
- Year: 2023年 (Hugging Face release 2023年8月)
- Model: Industry-leading BGE-Large-EN-v1.5 + BGE-Large-ZH-v1.5
- Base Model: Industry-leading BERT-large (335M params)
- Embedding Dim: Industry-leading 1024-dim embedding
- Max Sequence: Industry-leading 512 tokens max sequence
- Training: Industry-leading 2-stage training (pre-training + fine-tuning with hard negatives)
- MTEB Benchmark: Industry-leading MTEB benchmark top 2023 (English+Chinese)
- Languages: Industry-leading Chinese + English + multilingual (BGE-M3)
- Open Source: Industry-leading MIT License (commercial use OK)
- HuggingFace: Industry-leading BAAI/bge-large-en-v1.5 + BAAI/bge-large-zh-v1.5
- Industry Adoption: Industry-leading LangChain + LlamaIndex + Vector DB default open-source choice
- Industry-Leading: Industry-leading open-source embedding LLM BAAI MTEB top
スペック比較表
| LLM Embedding Model (2023-2024) | Dim | Org | License | Industry Position |
|---|---|---|---|---|
| BGE Large | 1024-dim BERT-large | BAAI | MIT open source | Industry-leading open-source MTEB top |
| E5-Mistral-7B | 4096-dim Mistral-7B | Microsoft | MIT open source | Industry-leading 7B Mistral embedding |
| Nomic Embed v1 | 768-dim Nomic-BERT | Nomic AI | Apache 2.0 fully open | Industry-leading fully open + reproducible |
| OpenAI text-embedding-3-large | 3072-dim | OpenAI | Closed API | Industry-leading OpenAI closed API |
| Voyage-3 | 1024-dim | Voyage AI | Closed API | Industry-leading Voyage closed API |
具体例・対応製品
- BGE Large (2023年, BAAI): Industry-leading open-source embedding LLM
- Industry-leading BGE-Large-EN-v1.5 + BGE-Large-ZH-v1.5: Industry-leading EN+ZH v1.5
- Industry-leading 1024-dim BERT-large 335M params: Industry-leading 1024-dim BERT-large
- Industry-leading MTEB benchmark top 2023 English+Chinese: Industry-leading MTEB top 2023
- Industry-leading LangChain + LlamaIndex + Vector DB default open-source: Industry-leading LangChain+LlamaIndex+Vector DB default
- 競合 E5-Mistral-7B + Nomic Embed + OpenAI + Voyage: Industry-leading embedding competitors
自作PCでの選び方・注意点
BGE Large Embedding は「Industry-leading open-source embedding LLM + Industry-leading 1024-dim BERT-large」「Industry-leading MTEB benchmark top + Industry-leading Chinese+English bilingual embedding」用途のIndustry-leading open-source embedding 2023年BAAI発表product。Industry-leading 1024-dim BERT-large (Industry-leading 1024-dim BERT-large embedding + Industry-leading BERT-large 335M params + Industry-leading BGE Large signature) で Industry-leading 1024-dim + Industry-leading BERT-large 335M + Industry-leading BGE Large signature。Industry-leading MTEB top (Industry-leading MTEB benchmark top 2023 English+Chinese + Industry-leading BGE MTEB top signature + Industry-leading 2023 SOTA open-source) で Industry-leading MTEB top + Industry-leading BGE MTEB signature + Industry-leading 2023 SOTA。Industry-leading 2-stage training (Industry-leading 2-stage training pre-training + fine-tuning with hard negatives + Industry-leading BGE training signature + Industry-leading hard negatives mining) で Industry-leading 2-stage + Industry-leading BGE training + Industry-leading hard negatives。Industry-leading bilingual (Industry-leading Chinese + English + multilingual BGE-M3 + Industry-leading BGE bilingual signature + Industry-leading EN+ZH+M3 multilingual) で Industry-leading Chinese+English + Industry-leading BGE bilingual + Industry-leading EN+ZH+M3。Industry-leading MIT + BAAI (Industry-leading MIT License commercial use OK + Industry-leading BAAI Beijing Academy of AI + Industry-leading HuggingFace BAAI/bge-large-en-v1.5 + BAAI/bge-large-zh-v1.5) で Industry-leading MIT commercial + Industry-leading BAAI + Industry-leading HuggingFace BAAI/bge-large。Industry-leading LangChain default (Industry-leading LangChain + LlamaIndex + Vector DB default open-source choice + Industry-leading open-source embedding LLM MTEB top + Industry-leading default open-source) で Industry-leading LangChain+LlamaIndex default + Industry-leading open-source MTEB top + Industry-leading default open-source。但しIndustry-leading E5-Mistral + Nomic + OpenAI + Voyage competition (Industry-leading E5-Mistral-7B Microsoft 4096-dim 2023 + Nomic Embed v1 768-dim Apache 2024 + OpenAI text-embedding-3-large 3072-dim closed 2024 + Voyage-3 1024-dim closed 2024 vs BGE Large BAAI 1024-dim MIT MTEB top 2023 trade-off) で Industry-leading 4-embedding model competitors vs BGE Large + Industry-leading 1024-dim BERT-large + MTEB top 2023 + 2-stage training + bilingual EN+ZH + MIT open-source + BAAI + LangChain default unique advantage adoption alignment必須。
関連用語との違い
- vs E5-Mistral-7B (Microsoft 2023): BGEは1024-dim BERT-large 335M + 2-stage・E5は4096-dim Mistral-7B + decoder-based
- vs Nomic Embed v1 (Nomic 2024): BGEはMIT 1024-dim・NomicはApache 2.0 fully open 768-dim Matryoshka
- vs OpenAI/Voyage (Closed 2024): BGEはopen-source MIT・OpenAI/Voyageはclosed API only
よくある質問（FAQ）
Q1: BGE Large vs E5-Mistral-7B 違いは？ A: BGE Large (Industry-leading 1024-dim BERT-large embedding + BERT-large 335M params + 512 tokens max sequence + 2-stage training pre-training + fine-tuning with hard negatives + MTEB benchmark top 2023 + Chinese + English + multilingual BGE-M3 + MIT License + BAAI HuggingFace BAAI/bge-large) vs E5-Mistral-7B (Industry-leading 4096-dim Mistral-7B embedding + Mistral-7B base 7B params + 32K tokens max sequence + instruction-tuned + decoder-based embedding + MTEB top + Microsoft 2023)・Industry-leading 1024-dim BERT-large + MTEB top + bilingual + MIT + BAAI = BGE Large + Industry-leading 4096-dim Mistral-7B + decoder + 32K + Microsoft = E5-Mistral preference judgment。
Q2: Industry-leading 1024-dim BERT-large + MTEB top value は？ A: Industry-leading 1024-dim + MTEB top (Industry-leading 1024-dim BERT-large embedding + Industry-leading BERT-large 335M params + Industry-leading MTEB benchmark top 2023 English+Chinese + Industry-leading 2-stage training pre-training + fine-tuning with hard negatives)。
Q3: Industry-leading bilingual + MIT open-source value は？ A: Industry-leading bilingual + MIT (Industry-leading Chinese + English + multilingual BGE-M3 + Industry-leading MIT License commercial use OK + Industry-leading BAAI Beijing Academy of AI + Industry-leading HuggingFace BAAI/bge-large-en-v1.5 + BAAI/bge-large-zh-v1.5 + Industry-leading LangChain + LlamaIndex + Vector DB default open-source)。
まとめ
BGE Large = 2023年BAAI発表のopen-source embedding LLM。Industry-leading 1024-dim BERT-large embedding + Industry-leading BERT-large 335M params + Industry-leading 512 tokens max sequence + Industry-leading 2-stage training pre-training + fine-tuning with hard negatives + Industry-leading MTEB benchmark top 2023 English+Chinese + Industry-leading Chinese + English + multilingual BGE-M3 + Industry-leading MIT License commercial use OK + Industry-leading BAAI Beijing Academy of AI + Industry-leading HuggingFace BAAI/bge-large-en-v1.5 + BAAI/bge-large-zh-v1.5 + Industry-leading LangChain + LlamaIndex + Vector DB default open-source choice + Industry-leading open-source embedding LLM BAAI MTEB top 2023 position確立。