Hugging Face Introduces Pooling-Aware Training for Multi-Vector Embeddings
Multi-vector embedding models are storage-intensive. Hugging Face presents pooling-aware training, which teaches encoders to preserve key information when token vectors are merged. With pool_factor=4, the trained model retains 97.8% of unpooled nDCG@10 versus 94.6% without training, at 4x storage reduction.