Report
CoBPE is claimed to produce 30% shorter English token sequences
A post says CoBPE uses six tokens where BPE uses 13 for a Hobbit line, without changing the text.
TLDR
A post comparing CoBPE with BPE says CoBPE attaches short words and punctuation to other words, reducing a quoted Hobbit line from 13 tokens to six without changing the text. It claims 30% shorter sequences across English and better LLMs at the same compute.
Combined views
3.5K
4 Sources, first seen ago