Reproduce “Learn to Pool: Lightweight Fine-Tuning for Flexible Multi-Vector Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Stefan Josef
The paper introduces a lightweight fine-tuning method for pre-trained ColBERT models to enable multi-vector compression while preserving or improving retrieval accuracy. This approach achieves substantial vector reduction as much as 83% on BEIR SciFact with no performance loss by integrating pooling strategies directly into the training process.
Key insights
- Pooling-aware fine-tuning is crucial for maintaining or improving retrieval accuracy with vector compression; training without pooling significantly degrades performance.
- The fine-tuning process makes token embeddings generally more amenable to aggregation, demonstrating positive transfer effects across different pooling methods and unseen datasets.
- Training with k-means pooling proves particularly effective, consistently yielding strong performance gains for vector compression across various datasets and pool factors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.