I Distilled a 397B Model Into a 4B One for $1.60. It Now Catches 28 of 33 Park Alerts That Should Stop Your Hike
Описание и применение
Модель Qwen3.5‑4B, обученная с помощью LoRA, повышает точность классификации уведомлений национальных парков до 74,6 % и выявляет 28 из 33 критических предупреждений, при этом стоимость запуска составляет $0,13 за 1 000 уведомлений.
Задача
National Park Service alerts are long and many hikers ignore them, leading to unsafe hikes.
Результат
The 4B LoRA model improves decision accuracy from 64.3% to 74.6% and catches 28 of 33 critical alerts, while the entire pipeline costs $1.60 to build and $0.13 per 1,000 alerts to run.
Рабочий процесс
- Collect alerts via NPS API for selected parks
- Run each alert through fine‑tuned Qwen3.5‑4B LoRA
- Generate a printable card with overall verdict and trail list
- Print or screenshot the card
- Use card to decide whether to hike
Доказательства
On 126 held‑out alerts the fine‑tuned 4B went from catching 17 of 33 alerts that should stop a hiker to 28 of 33, and from 64.3% to 74.6% decision accuracy.
It costs $0.13 per 1,000 alerts, about 11× less than the teacher, and the whole ML pipeline cost $1.60.
The model's only job is to produce one printable card a week, so you can stop checking park alerts and go outside.
- Дата записи: 2026-10-10T14:56:20.393Z