Finetuning Qwen3-VL-Embedding boosts Visual Document Retrieval accuracy
tom aarsen released a guide on finetuning multimodal embedding models using sentence transformers. by finetuning qwen/qwen3-vl-embedding-2b on a visual document retrieval task, the model achieved an ndcg@10 of 0.947 compared to the base model's 0.888. this custom model outperformed existing vdr models up to 4x its size.
Hugging Face technical lead Tom Aarsen published a guide on training and finetuning multimodal embedding and reranker models using Sentence Transformers. The tutorial demonstrates a Visual Document Retrieval (VDR) workflow, matching text queries to document page images. Aarsen finetuned the Qwen/Qwen3-VL-Embedding-2B model, improving its NDCG@10 score from 0.888 to 0.947 on evaluation data. This performance gain exceeds larger, existing VDR models (up to 4x the parameter count), highlighting the efficiency of task-specific adaptation.