Skip to content

Model development#

Train models#

  1. Install dependencies: uv sync (CUDA 13.0 is supported out-of-the-box)
  2. Fetch datafiles: uv run weasel assets
  3. Build all the models: GPU=1 uv run weasel run all

Fine-tune the models#

Hyperparameters of the underlying models can be fine-tuned using Weights&Biases: wandb sweep with one of the sweep_*.yml config file.

Publish models#

  1. Make sure dependencies are up-to-date
  2. Bump version: uv run bumpversion --new-version x.x.x major/micro/patch --verbose
  3. Build the model as described in the previous section
  4. Publish the new model to Hugging Face Hub: uv run weasel run publish (must be executed in the model's directory)