SwanLab

An open-source Chinese experiment-tracking and visualization tool for model training, similar to W&B.

SwanLab is an open-source experiment-tracking tool built by a team in China, functionally similar to Weights & Biases: add a few lines to training code, and it logs loss curves, hyperparameters, images, and hardware usage to a web dashboard, making it easy to compare multiple runs and share results with a team. It offers both a cloud-hosted version and a self-hostable one, is directly accessible from within mainland China, and integrates with common frameworks such as PyTorch, Hugging Face Transformers, and LLaMA-Factory. It's a common substitute when training models in China, where reaching W&B's servers can be inconvenient.

ExampleCall swanlab.init in a training script, then swanlab.log at every step to record the loss, and compare the curves of three different learning rates on the web dashboard.

Related

Weights & Biases (W&B) · TensorBoard · Hyperparameter · LLaMA-Factory / ms-swift (ModelScope SWIFT)

Sources

SwanHubX/SwanLab (GitHub)