Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Sentence Transformers v6.0 adds support for training ColBERT-style multi-vector models, which match queries to documents token-by-token instead of using a single compressed vector, preserving fine-grained signals that improve retrieval accuracy. A model finetuned using this approach on medical data outperformed all general-purpose retrievers by +0.062 NDCG@10, trained in 14.5 hours on a single consumer GPU.