preprint
Open access
Contrastive Geometric Transfer: Efficient Sentence Embeddings via Hyperbolic Projection with 24× Compression
Research footprint
At a glance
- Citations
- 0
- References
- 0
- Comments
- 0
Paper overview
Abstract
Experimental results for Contrastive Geometric Transfer (CGT), a method for compressing sentence embeddings via hyperbolic projection. Achieves 24× compression (768d→32d) with 97.4% quality retention (Spearman ρ=0.812 vs baseline 0.834). Outperforms PCA by 5.9% at 32d and 16.7% at 16d. Identifies semantic degradation threshold at 5 dimensions. Evaluated on STS-B benchmark using MPNet and MiniLM teacher models. Includes LaTeX manuscript, JSON experimental results, and Jupyter notebooks for reproducibility
Record transparency
Publication details
- DOI
- 10.5281/zenodo.18225189
- OpenAlex
- W7123244425
- Document type
- preprint
- Language
- EN
- Source
- Zenodo (CERN European Organization for Nuclear Research)
- Last metadata update
Comments
Log in to join the discussion.