Remix.run Logo
imperio59 13 hours ago

Pre-training data is pre-tokenized ahead of time before being used to not waste any GPU compute.

A massive speedup like this is a nice efficiency savings on some of these data pipelines for sure.