Training Studio 0.1.3

Training Studio 0.1.3

Training Studio has a simpler interface and a continuous loop for local model training.
This release replaces the six-hour limit for new Studio sessions.
Existing sessions keep their saved results and controls.

Interface

• A new logo identifies the app and its Windows shortcuts.
• The Mica background and translucent panels use the Windows transparency settings.
• The appearance defaults to the Windows theme. Light and Dark remain available.
• One main button changes between Start and Stop. A separate Pause button releases the GPU for gaming.
• Icons and text show the current state, including Running, Paused, Stopped, and Needs attention.
• Home, Activity, Results, and Settings separate daily controls from setup details.
• Background updates no longer disable the toolbar or cause its controls to blink.

Training loop

• New sessions collect code, train a candidate, compare results, and repeat until Pause or Stop.
• The recipe uses Qwen3.5-4B with Python, Go, TypeScript, Rust, and Dart examples.
• The loop keeps generated programs only after they pass code tests in isolated containers.
• Earlier passing examples remain in the training data to reduce loss of previous skills.
• The base weights stay frozen. Small LoRA adapters receive at most ten optimizer steps per cycle.
• A candidate must gain a validation pass without losing a previous pass before the full comparison.
• The full comparison covers 790 tasks. It also runs every fifth cycle that has a training candidate.
• A candidate replaces the retained model only after the required comparisons pass without lost tasks.
• A rejected candidate does not end the loop. The loop keeps its report and collects new examples.
• Pause preserves progress. Start resumes it. Stop ends the session and preserves its results.
• Login recovery respects saved Pause and Stop commands.

Install or update

1. Download TrainingStudio-win-x64.zip from this release.
2. Extract all files into one folder.
3. Run install.ps1 with PowerShell.
4. Open Training Studio from the Desktop or Start menu.

The package includes .NET and its license notices. Windows can show an unknown-publisher prompt because the executable has no code signature.
The installer keeps an active worker on its current version. Shortcuts and future worker starts use the new version.

If an older dashboard remains open, close it before you open the new shortcut.
Keep the older app folder until its worker exits.
If an experiment remains active, keep its Python source files unchanged.
Changed source files can prevent that experiment from resuming.

Requirements and limits

The app requires a prepared WSL CUDA backend, model cache, dataset cache, and isolated code runtime.
It does not install these dependencies. The current recipe targets a 16 GB NVIDIA GPU.

The new loop does not yet establish better coding quality. Public benchmark scores are development results because the loop reuses these tests.
Model files, private data, checkpoints, and local reports are not part of this release.

Installation guide · Learning method

Training Studio 0.1.2 — GPU power controls

Training Studio now includes a GPU power slider and 5% presets within the NVIDIA driver's supported range. The app shows the requested percentage, wattage and actual applied limit.

Pause restores 100%; Resume reapplies the saved training limit. Stop, failure and completion restore 100%. These controls work with the dashboard closed. Limits are saved per experiment, and the current model and training recipe are unchanged.

A one-time administrator setup installs protected, fixed NVIDIA preset actions. The app and learning worker keep normal user permissions. Power limits affect all apps on the selected GPU. If the Windows worker is force-killed, use Restore 100% if its limit remains.

Verified on Windows hardware: 85/90/95/100% presets, normal-user execution with command editing denied, UI slider input, Pause/Resume with the dashboard closed, preserved paused clock, and native UI rendering. The preview requires a prepared WSL CUDA backend and is not code-signed.

Training Studio 0.1.1 — startup preferences

Training Studio now has independent startup switches for learning and the dashboard. Manual mode removes automatic login, timer and failure-retry starts. Saved Pause and Stop stay respected. Changing startup preferences leaves current work running.

The installer keeps new versions beside an active worker and updates future launches without interrupting training.

Validation: native Windows checks of all four startup combinations, live dashboard rendering, Windows build, and bridge/controller tests. This preview needs a prepared WSL CUDA backend. The executable is not code-signed.

Training Studio 0.1.0 — Windows preview

Training Studio is a native C# Windows app for local CUDA fine-tuning through WSL.

• Live activity, GPU use, saved checkpoints and readable worker logs.
• Matched coding results with visible gained and lost tasks, plus experiment history.
• Start, Pause for gaming, Resume and Stop. Saved controls survive Windows login.
• A wizard that checks private training data and prepares a paused Qwen3.5-4B experiment.

Download and extract TrainingStudio-win-x64.zip. Run install.ps1, or launch TrainingStudio.exe from a permanent folder. The release includes .NET and license notices.

This first release requires a prepared WSL CUDA backend. The wizard supports Qwen3.5-4B Go/TypeScript repair training. It does not install CUDA, download models, or support arbitrary architectures yet. The executable is unsigned.

Validation: 22 bridge/controller tests on Linux, native Windows input checks, all five screens rendered on Windows, and a Windows/WSL control test covering scheduler launch, a fixed paused clock, resume, and stopped-session recovery. A real CUDA experiment was displayed during training. These checks do not establish model quality or test a new-machine installation from scratch.

Source and setup: https://github.com/M1chlCZ/local-coding-assistant/tree/main/desktop