Run on Your Local Workstation

When launching examples locally, uv will automatically download and keep your local environment in sync with the package’s requirements.

For example, to finetune a small Qwen3 model on the HellaSwag dataset, simply run:

uv run recipes/llm/finetune.py --model.pretrained_model_name_or_path Qwen/Qwen3-0.6B

To finetune a slightly larger model on multiple GPUs and sharded using FSDP2, you can make a slight augmentation to the above command. For example, on 2 GPUs simply run:

uv run torchrun --nproc-per-node=2 recipes/llm/finetune.py --model.pretrained_model_name_or_path Qwen/Qwen3-1.7B

finetune.py uses the default config file. You can easily customize the config by passing in command-line arguments, editing the config directly, or creating your own configuration file and passing it using --config /path/to/config.