Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers

The Sentence Transformers library enables the training and finetuning of multimodal embedding and reranker models that handle text, images, and other data types. As a practical example, the process of finetuning a Qwen3-VL-Embedding-2B model for Visual Document Retrieval (VDR) is detailed, which involves matching text queries to document page images. The training pipeline requires a model, a dataset, a loss function, and the SentenceTransformerTrainer to bring the components together. Finetuning on domain-specific data is shown to significantly boost performance, with the example model's NDCG@10 score improving from 0.888 to 0.947.