Volunteer setup (Windows)#

A volunteer lends their own GPU to a coordinator. Training happens locally; only the small LoRA adapter (~40 MB) and the data shard served by the coordinator cross the network.

Dikkat

Status: the install script is written but has not yet been tested end to end on a clean machine, and the release packages (modelen.exe, modelen-istemci.zip) will be produced with the first tagged release. Try the script at your own risk, with -Dene first; report problems on GitHub.

Requirements#

Windows 10/11 (64-bit)
An NVIDIA GPU with at least 6 GB VRAM (tested with 8 GB) and a current NVIDIA driver
~8 GB free disk (PyTorch about 3 GB, base model 1.2 GB)
An internet connection (downloads are hundreds of MB; mind metered connections)
Administrator rights are not needed

Install#

Download kur.ps1 from the repository. First see what it would do (installs nothing):

powershell -ExecutionPolicy Bypass -File kur.ps1 -Dene

Then install:

powershell -ExecutionPolicy Bypass -File kur.ps1

Parameter	Meaning
`-Hedef`	Install folder (default `%LOCALAPPDATA%\Modelen`)
`-Surum`	GitHub release tag (default: latest)
`-CudaSurumu`	PyTorch wheel: `cu124` (default), `cu121`, `cu126`, `cpu` (testing only)
`-Kaynak`	Offline/local install: the repository root (containing `istemci\`)
`-ModelAtla`	Skip the base model download
`-Dene`	Only check prerequisites and print the plan

What gets installed?#

Everything lives in one folder (%LOCALAPPDATA%\Modelen); your system Python is untouched.

Path	Content
`modelen.exe`	Desktop app (Easy mode: become a worker with one button)
`istemci\`	Worker scripts
`.venv\`	Python 3.12 + PyTorch + `transformers`, `peft`, `safetensors` (installed with uv)
`models\qwen3-0.6b-base\`	Base model (Qwen3-0.6B-Base, Apache-2.0)
`arac\uv.exe`	Install tool

The script compares the SHA-256 of release assets with SHA256SUMS. A Modelen shortcut is added to the Start menu.

Use#

Open Modelen from the Start menu (Easy mode).
Choose a coordinator (approved directory). The coordinator decides which data and mode to train with; read its description before choosing.
Press ON. The app takes a task, trains and uploads the result. Press OFF to stop.

The worker downloads the coordinator's data shard in the first round and caches it under %USERPROFILE%\.modelen\veri; later rounds do not download it again.

Privacy and security#

Training happens on your GPU; only the trained adapter goes to the coordinator. No personal data other than the pseudonym is sent.
Weights and data from the network are read only as safetensors (no pickle, no code execution), and their digests are checked.
Think about who a coordinator is before trusting it: it chooses the training data. See Security.

Removal#

Delete the %LOCALAPPDATA%\Modelen folder and the Modelen.lnk shortcut in the Start menu. Settings live in %APPDATA%\Modelen (API keys are in the Windows credential store).