Training and evaluating an assistant that runs offline
Fine-tuning local
Alban Pasquelin built a preparation and evaluation harness to specialise a multilingual Qwen model driving AI Desktop Studio, run locally on Apple Silicon inside Tart virtual machines.
import { Qwen, AppleSilicon, Tart, Nodejs } from '@/stack'
export function specialiseModel(base: Qwen): LocalAssistant {
A general model cannot drive a specific application. This one is specialised to understand the 310 actions of AI Desktop Studio, across several languages.
const model = qwen.specialise({ actions: 310, languages: 'multi' })
Everything happens on the machine: data preparation, training, then an automated evaluation measuring whether the new version genuinely beats the old one. Without that measurement you are not fine-tuning, you are practising superstition.
if (evaluate(candidate) <= evaluate(shipped)) reject(candidate)
return {
runLocally: 100%, /** run locally **/
}
}