The 23-Minute Fine-Tune Recipe — train a private AI on the computer you already own, and beat the best prompt you ever wrote

The companion to the video. On August 19, 2026 we fine-tuned a 4-billion-parameter model (Qwen3-4B) on an ordinary 24 GB MacBook with Unsloth Studio — 540 real customer-support examples, 22 minutes 48 seconds of training, zero dollars. Then it took an exam of 54 messages it had never seen: 98.1% correct, with no answer list in its prompt — while the very same base model, handed the complete 27-category answer key AND worked examples, managed only 74.1% zero-shot and 79.6% few-shot. The model holding the answer key lost. INSIDE: (1) THE 5-STEP RECIPE — the one-line install, model choice, writing 200–500 examples from tickets and emails you already have, the sacred hold-out exam, and training on defaults. (2) THE MACHINE-FIT TABLE — what a 3B/4B/7B/8B fine-tune actually needs (3.5–6 GB), and which Macs and gaming GPUs qualify. (3) THE 5 SETTINGS THAT MATTER — including target_modules, the setting whose absence killed our first run with a cryptic '[grad]' error. (4) COPY-PASTE DATA TEMPLATES — the exact one-line-per-example chatml format Studio auto-detects, adapted for routing, brand voice, and extraction. (5) THE FULL RECEIPTS — our complete before-and-after test, the one question it missed (shown, not hidden), and the honest rules for rerunning it on your own numbers, plus when fine-tuning is the WRONG tool and retrieval wins. Independent playbook from Hyperautomation Labs — not affiliated with Unsloth AI or Bitext.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.