Want to adapt Cosmos 3 to your own video data without spending days manually configuring training, evaluation, and hyperparameter sweeps?
Whether you’re exploring post-training or already testing a dataset, these resources show how to:
- Establish a zero-shot baseline before training.
- Run LoRA post-training with a coding agent and NVIDIA TAO agent skills.
- Use TAO AutoML to optimize the configuration and compare results.
On the Woven Traffic Safety dataset, exact-match video Q&A accuracy increased from 54.41% at baseline to 87.14% with LoRA and 93.35% with AutoML. Results will vary by task and dataset.
Choose how you want to get started:
📖 Read the Technical Blog
📺 Watch the Step-by-Step Tutorial
🧑💻 Join NVIDIA Experts for the Livestream
🛠️ Join us for Cosmos Labs Office Hours