Readme

Supermix_29

Supermix_29 is a mixed research and packaging workspace for local AI experiments. It combines:

• Supermix / ChampionNet model research assets
• Qwen-based LoRA training pipelines
• a packaged local Python chat runtime
• a Windows desktop chat app build pipeline
• a browser-only metadata chat for GitHub Pages

This repository is not source-only. It also contains datasets, training artifacts, logs, packaged outputs, and installer files.

Naming note

The local folder and origin remote use Supermix_27, while some branches, documents, and artifact names refer to Supermix_28 or Supermix_29. Treat those as experiment-line and snapshot names inside the same evolving project.

Repository layout

.
|-- source/                Main development workspace
|-- runtime_python/        Packaged local inference runtime
|-- web_static/            Browser-only metadata chat assets
|-- installer/             Inno Setup assets
|-- artifacts/             Training outputs, adapters, checkpoints
|-- datasets/              Training data inputs
|-- dist/                  Built desktop outputs
|-- build/                 Build staging/output
|-- assets/                Branding and packaging assets
|-- ARCHITECTURE.md        Architecture notes
|-- MODEL_CARD_V28.md      Model card for the v28 line
`-- README.md

Main entrypoints

Use runtime_python/ if you want the fastest way to run the packaged local runtime.

Use source/ if you want the current development scripts for training, chat app work, desktop packaging, and training-monitor automation.

Key files:

• source/qwen_supermix_pipeline.py: main Qwen training pipeline
• source/qwen_chat_web_app.py: current local web chat app
• source/qwen_chat_desktop_app.py: desktop app entrypoint used by PyInstaller
• source/training_monitor_gui.py: GUI monitor for active training runs
• runtime_python/chat_web_app.py: packaged runtime web app
• web_static/index.html: browser-only metadata chat UI

Prerequisites

Runtime dependencies:

python -m pip install -r runtime_python/requirements_runtime_interface.txt

Training and build dependencies:

python -m pip install -r source/requirements_train_build.txt

Typical extras for desktop packaging:

python -m pip install pywebview pillow pyinstaller

Notes:

• Windows is the primary platform for the desktop and training-automation workflows.
• The Qwen pipelines expect a local or cached Qwen/Qwen2.5-0.5B-Instruct base model unless you override the default.
• Optional accelerators include CUDA and DirectML, with fallback to CPU.

Quick start

Run the packaged local runtime

python runtime_python/chat_web_app.py

Windows launchers are also included:

runtime_python\launch_chat_web_supermix.bat
runtime_python\launch_chat_terminal_supermix.bat

This path is the best fit if you want to run the packaged checkpoint and metadata bundle without using the full development workspace.

Run the current source web app

python source/qwen_chat_web_app.py

Use this path when you want the current development version that works with the latest adapter artifacts and desktop packaging flow.

Run the browser-only metadata chat

Open web_static/index.html in a browser and load:

web_static/chat_model_meta_supermix_v27_500k.browser.json

Important limitation: this is metadata-driven browser chat, not full PyTorch inference in the browser.

Training workflows

The current training scripts are centered on source/qwen_supermix_pipeline.py and the v28 recipe family.

Smoke run

powershell -ExecutionPolicy Bypass -File run_train_qwen_supermix_v28_smoke.ps1

This writes a short validation run under:

artifacts\qwen_supermix_enhanced_v28_improvements_smoke

Full auto-resume run

The repo includes a Windows launcher that starts the latest full recipe and reattaches to the newest checkpoint when possible:

launch_train_qwen_supermix_v26_full.bat

Equivalent direct command:

powershell -ExecutionPolicy Bypass -File source\auto_resume_supermix_training.ps1

By default this targets:

artifacts\qwen_supermix_enhanced_v28_clean_eval_robust_ipo

and warm-starts from:

artifacts\qwen_supermix_enhanced_v26_full

Training monitor

Start the monitor GUI with:

source\launch_training_monitor_gui.bat

or:

python source/training_monitor_gui.py --root .

The monitor parses run logs, reports stage progress, and surfaces runtime/device details for active training jobs.

Register auto-resume at login

powershell -ExecutionPolicy Bypass -File source\register_supermix_auto_resume_task.ps1

This tries to create a scheduled task and falls back to an HKCU\Software\Microsoft\Windows\CurrentVersion\Run entry when scheduled tasks are unavailable.

Desktop packaging

Build the desktop application:

powershell -ExecutionPolicy Bypass -File build_qwen_chat_desktop_exe.ps1

Expected output:

dist\SupermixQwenDesktop\SupermixQwenDesktop.exe

The build script generates branding, resolves the latest adapter artifact automatically, stages a desktop bundle, and packages the app with PyInstaller.

Build the installer:

powershell -ExecutionPolicy Bypass -File build_qwen_chat_desktop_installer.ps1

Expected output:

dist\installer\

The installer flow requires Inno Setup 6. If iscc.exe is not available, install it with:

winget install --id JRSoftware.InnoSetup -e --accept-package-agreements --accept-source-agreements

Tests

This repo includes direct-run smoke and regression tests for the chat app, training pipeline, monitor, and expert variants.

Examples:

python test_qwen_chat_web_app.py
python test_training_monitor_gui.py
python test_training_resume_automation.py

Additional experiment-specific tests live both at the repository root and under source/.

Important docs

• ARCHITECTURE.md
• MODEL_CARD_V28.md
• source/CHAT_FINETUNE.md
• source/RESEARCH_UPGRADES.md

Limitations and repo shape

• This repo contains generated artifacts and logs alongside source code.
• Naming is mixed across Supermix_27, Supermix_28, and Supermix_29.
• The browser build is metadata-only and does not run the full model in-browser.
• Desktop packaging and training automation are Windows-first.
• Some flows assume a locally available base model and local Python environment.

License

See LICENSE.