Model development#
Train models#
- Install dependencies:
uv sync(CUDA 13.0 is supported out-of-the-box) - Fetch datafiles:
uv run weasel assets - 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#
- Make sure dependencies are up-to-date
- Bump version:
uv run bumpversion --new-version x.x.x major/micro/patch --verbose - Build the model as described in the previous section
- Publish the new model to Hugging Face Hub:
uv run weasel run publish(must be executed in the model's directory)