Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Hugging Face has announced the release of Sentence Transformers v6.0, featuring a new MultiVectorEncoder model type designed for ColBERT-style late interaction retrieval. According to the source, multi-vector models preserve one small vector per token rather than compressing an entire text into a single summary vector. This token-level matching captures fine-grained domain signals that traditional dense models average away, resulting in stronger retrieval performance. Training a multi-vector model involves components such as datasets, loss functions, training arguments, evaluators, and the MultiVectorEncoderTrainer class. The update enables developers to train or finetune models from scratch or from existing checkpoints, handling custom document length configurations that prevent the truncation issues common in older models. In evaluations, a medical retrieval model trained on a single RTX 3090 over 14.5 hours outperformed general-purpose retrieval models. Experiments also showed that unsupervised checkpoints adapt to new domains significantly better than finished supervised checkpoints, while fresh projections on retrieval-pretrained backbones serve as strong alternatives. The library supports loading training and evaluation data from local formats or the Hugging Face Datasets Hub, operating with pip install -U sentence-transformers train.