SettledResearch
Sentence Transformers adds pooling-aware training for multi-vector retrieval
A Nov. 11 Hugging Face post by Tom Aarsen describes token-pooling-aware training for multi-vector embedding models. On a ModernBERT-base model, pooling by a factor of 4 cuts vectors per document 74.8% and keeps 97.8% of the unpooled NanoBEIR score, versus 94.6% with ordinary training. Code and checkpoints are released.
- HEAT
- OUTLETS
- 0
- FIRST SEEN
- LAST SIGNAL
Primary source
Hugging Facehuggingface.co/blog/token-poolingCoverage · 1 article, oldest first
Tuesday, November 10, 2026
Hugging FacePrimary
Token Pooling-Aware Training for Multi-Vector Embedding Models