Reproduce “Luxical: High-Speed Lexical-Dense Text Embeddings” 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
DatologyAI, Luke Merrick, Alex Fang, Aldo Carranza, Alvin Deng, Amro Abbas, Brett Larsen, Cody Blakeney and 25 more
Luxical converts sparse n-gram counts into reusable dense document embeddings with a small neural network trained to match a teacher’s pairwise similarities. In a 10% FineWeb filtering experiment, its scoring pipeline processed 23.1 MiB/s, versus 1.6 MiB/s for FineWeb-Edu, while resulting 3B-parameter models had similar accuracy, including 36.4% MMLU for both. Results cover English settings; top-1 retrieval remains weaker.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.