Hayate
Hayate 疾風
Distilling a big LLM into a web-page categorizer that answers in under a millisecond, one weekend at a time.
One glance, one cut. Your page is tagged before you blink.
Progress log
Weekend 3: who teaches Hayate-chan?
Picking the LLM that labels my training data: three prompts, four models, a fix that went too far, and a consistency number that was too good to be true.
Weekend 2: twenty thousand pages from Common Crawl
Sampling Portuguese and English pages fairly from a 150 GB index without downloading it, and getting blocked once along the way.
Weekend 1: pinning down 330 labels
Setting up the repo, downloading a taxonomy that keeps changing under its version number, and deciding how deep the labels go.