Unverified50% confidenceFactExact time
EmbeddingGemma2基于Gemma 2语言模型架构，经过多模态对比学习训练
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
Unverified专用embedding模型通常基于双编码器（Bi-Encoder）架构训练，通过对比学习优化问题-段落向量距离，这是检索专用模型召回率优于通用模型的根本原因73% similarUnverifiedEmbeddingGemma2采用Matryoshka表示学习，将最重要的信息编码到向量最前面，支持向量截断70% similarUnverifiedEmbedding模型本质上是Encoder-only Transformer架构，代表为BERT及其变体，通过对比学习或掩码语言模型任务预训练70% similarUnverifiedEmbedding模型微调通常采用对比学习框架，通过构造正样本对和负样本对训练模型拉近正样本、推开负样本68% similarUnverified现代多模态大模型采用统一的端到端训练框架，图像编码器与语言模型联合优化67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/985549API
curl https://kongchang.com/api/v1/knowledge/claims/985549MCP
get_claim(id=985549)