About

Give Hayate a web page and she tells you what it's about: a marathon plan becomesSports › Running, a Lisbon hotel review becomes Travel › Hotels. The catch is that she has to do it in under a millisecond, on a plain CPU, for almost no money.

Pipeline: public pages, a big LLM labels them, tiny models learn to imitate it, then accuracy, speed and cost are measured

A big LLM (the teacher) labels public web pages once, using the IAB content categories. Small models (the students) learn to copy those labels, from hashed word patterns to a compact transformer. Each is measured for quality, latency and cost per million pages. The result is a trade-off curve, not a single model.

It's a personal weekend project that uses only public data and public models. It categorizes pages, not people.

Meet Hayate-chan

Hayate-chan is the model itself, drawn as a character (疾風, "gale"): smug, fast and slightly impatient with anything slow. When she gets a page right, that's the model; when she trips over a cookie wall, that's the model too. Her look is fixed in a canon, so she stays the same from post to post.