● Runs on your device · Tether QVAC

Pocket Tune

Teach a small language model your voice from your notes, a WhatsApp export, an FAQ or answers you speak out loud. It scrubs secrets, fine-tunes a LoRA adapter on your own computer or phone, and compares it with the base model. Your data never leaves the device.

Pocket Tune trains models, so it runs on your machine rather than in this page: QVAC needs your CPU or GPU and downloads a base model (Qwen3 0.6B is 382 MB).
Pocket Tune's web UI comparing the base model's answer with the tuned model's answer to the same prompt

What it does

Your data in

Markdown notes, WhatsApp exports, chat JSON, Q&A CSVs, or questions answered out loud and transcribed by Whisper on the device.

Scrubbed first

Emails, phone numbers, wallet addresses, private keys, tokens, IBANs, cards and seed phrases are replaced before anything is trained.

Trained locally

QVAC fine-tunes a LoRA adapter with a memory check first, a live loss curve, and pause and resume.

Shared peer-to-peer

Publish the adapter (never the data) as a pear:// link with qvac-seed; a friend's Pocket Tune downloads it, checks its SHA-256 and compares it.

On a phone

Mockups of the Pocket Tune phone app: first-run guide, answering questions out loud, checking the transcript, training, and comparing
The Expo app: the same flow on an iPhone or Android phone (mockups of its real screens). Training runs on the CPU on iOS and the GPU on Android.

Sharing an adapter, no server

The Share tab publishing the pirate-voice adapter, seeding to one peer
Another Pocket Tune looking up the shared pear:// link before downloading
Two Pocket Tunes on one Mac over the public Hyperswarm network: looking up the link took about 2 s, and downloading and checking the 20 MB adapter 0.3 s.

Run it

# the web UI (Node.js 22+)
git clone https://github.com/lopeselio/qvac-pocket-tune
cd qvac-pocket-tune && npm install && npm run build:ui
npm start                     # opens http://127.0.0.1:4322

# or the CLI, from the same clone
node bin/pocket-tune.js prep examples/pirate-voice/faq.csv --out runs/pirate
node bin/pocket-tune.js train runs/pirate

The phone app is in mobile/ (Expo; build it with Xcode or Android Studio).