pyinstaller-demo/
├── scripts/
│   ├── 00_install_git.sh
│   ├── 01_install_miniconda.sh
│   ├── 02_create_conda_env.sh
│   ├── 03_install_uv.sh
│   ├── 04_git_init_project.sh
│   ├── 05_uv_sync.sh
│   ├── 06_run_active_learning_demo.sh
│   ├── 07_build_binary.sh
│   └── util_detect_os.sh
├── pyproject.toml
├── uv.lock                 # generated by uv
├── config/
│   └── extra-packages.txt   # future add‑on packages (one per line, with optional ==version)
├── src/
│   ├── __main__.py
│   ├── active_learning.py
│   ├── data_utils.py
│   └── model_utils.py
├── app.spec                # PyInstaller spec (generated by the build script)
├── README.md
└── .gitignore

The active-learn command runs an uncertainty sampling loop and fine-tunes syvai/speaker-diarization-3.1 as new segments are labeled. Example:

python -m src.__main__ active-learn --iterations 5 --query_k 3 \
  --lr 3e-5 --batch_size 2 --ft_epochs 2 \
  --ls_token YOUR_TOKEN --project_id 1 \
  --model_id syvai/speaker-diarization-3.1

--ls_token and --project_id configure the Label Studio project used for human-in-the-loop annotation. If omitted, a simple CLI simulator is used. --ls_url controls the Label Studio server address.

During each iteration the script fine-tunes syvai/speaker-diarization-3.1 on all labeled segments using mini‑batch updates. Use --batch_size, --lr and --ft_epochs to control this process. --model_id can override the base model (defaults to syvai/speaker-diarization-3.1).

Execute scripts/04_run_demos.sh to run the uncertainty-sampling loop over a small audio pool and save the model checkpoint.