The AIS研究院 recently released three new vector models, which have demonstrated outstanding performance in vector retrieval tasks and set new benchmarks in multiple evaluations. These models are:
BGE-EN-ICL: An English vector model that enhances semantic expression by introducing task-related query-document examples as few-shot learning instances.
BGE-Multilingual-Gemma2: A multilingual vector model that excels, particularly in enhancing Chinese and English capabilities.
BGE-Reranker-v2.5-Gemma2-Lightweight: A multilingual reranking model that supports layer-wise early output and token compression through optimized design, saving computational resources.
These models are trained on large language models, possessing excellent domain adaptation capabilities and extensive generalization performance. They also incorporate contextual learning and distillation techniques to enhance overall model performance and capabilities in retrieval tasks. The BGE-Reranker-v2.5-Gemma2-Lightweight model in particular focuses on lightweight design, making it efficient while maintaining superior performance.
In experimental results, these models have performed well in multiple evaluation benchmarks such as MTEB, BEIR, and AIR-Bench. BGE-Multilingual-Gemma2 stands out in multilingual capabilities, especially in enhancing Chinese and English abilities. BGE-EN-ICL is particularly notable for its few-shot performance. BGE-Reranker-v2.5-Gemma2-Lightweight has also achieved better results in reranking tasks and ensured superior performance while saving computational resources.
Model Links
（1） BGE-EN-ICL:
https://huggingface.co/BAAI/bge-en-icl
（2） BGE-Multilingual-Gemma2:
https://huggingface.co/BAAI/bge-multilingual-gemma2
（3） BGE-Reranker-v2.5-Gemma2-Lightweight:
https://huggingface.co/BAAI/bge-reranker-v2.5-gemma2-lightweight