Monitors¶
LLaMA-Factory supports multiple training visualization tools, including: LlamaBoard, SwanLab, TensorBoard, Wandb, MLflow.
LlamaBoard¶
LlamaBoard refers to the built-in Loss curve dashboard in the WebUI, which allows convenient viewing of Loss changes during the training process.
If you want to use LlamaBoard, simply start training using the WebUI.
SwanLab¶
SwanLab is an open-source training tracking and visualization tool that can be used both online and offline, supporting features such as hyperparameter logging, metric recording, multi-experiment comparison, hardware monitoring, and experiment environment recording, effectively helping developers manage experiments.
If you want to use SwanLab, please add the following parameters to the training configuration file when starting training:
use_swanlab: true
swanlab_project: llamafactory
swanlab_run_name: test_run
Alternatively, enable SwanLab recording in the SwanLab module of the WebUI:
Visualization example:
TensorBoard¶
TensorBoard is an open-source offline training tracking tool from TensorFlow, which can be used for recording and visualizing the training process.
If you want to use TensorBoard, please add the following parameters to the training configuration file when starting training:
report_to: tensorboard
Alternatively, enable TensorBoard recording in the Enable External Logging Panel under the Other Parameter Settings module of the WebUI:
Wandb¶
Wandb (Weights and Biases) is a cloud-based training tracking tool that can be used for recording and visualizing the training process.
If you want to use Wandb, please add the following parameters to the training configuration file when starting training:
report_to: wandb
run_name: test_run
Alternatively, enable Wandb recording in the Enable External Logging Panel under the Other Parameter Settings module of the WebUI:
MLflow¶
MLflow is an open-source offline training tracking tool from Databricks, used for recording and visualizing the training process.
If you want to use MLflow, please add the following parameters to the training configuration file when starting training:
report_to: mlflow
Alternatively, enable MLflow recording in the Enable External Logging Panel under the Other Parameter Settings module of the WebUI:
