Latent Action Model Training

February 14, 2026 · View on GitHub

The provided configuration only includes the GR-1 post-training dataset as an example. You can add new datasets by adding folders to dataset_paths by yourselves. The code will recursively search the folder and add any detected MP4 videos to the training set.

Using the following bash script to launch latent action training (suppose you are training on 1 node with 8 GPUs).

cd external/lam_project
bash train.sh

A multi-node training script example is provided at external/lam_project/launch.sh. Add the index of the YAML file to specify the configuration to use.

bash launch.sh 0

Note

Our latent action model weights can be found at Hugging Face. To set the path to your own checkpoint for proxy action extraction, check https://github.com/NVIDIA/DreamDojo/blob/main/cosmos_predict2/_src/predict2/models/text2world_model_rectified_flow.py.

<= Previous: [Setup]

=> Next: [DreamDojo Pretraining]