Post-Train NVIDIA Cosmos 3 in One Day Using Agent Skills

NVIDIA's Cosmos 3 experiment shows how task-specific training can improve answers about traffic videos. A coding agent used TAO workflow instructions to prepare data, evaluate the starting model, and train small adapter weights for its reasoning component. Accuracy on four-choice questions from the Woven Traffic Safety dataset rose from 54.41% to 87.14% after a run taking about 30 minutes on eight A100 GPUs. The reported experiment reached 93.35% after a much larger configuration search: 43 trials over 19.5 hours across five eight-GPU nodes. These are validation-set question-answering results; they do not measure driving performance.

“However, since SFT updates all model weights, it requires extensive computing resources and may show regression on general knowledge.”

Tanya Lenz

“For the example detailed in this post, LoRA post-training required ~7x fewer GPU hours compared to full-parameter SFT for Cosmos 3 Nano, making a one-day post-training turnaround a reality for engineering teams.”

Tanya Lenz

“In just one run, and around 30 minutes of training time on eight NVIDIA A100 Tensor Core GPUs, the model jumps to 87.14% accuracy, a massive improvement of 32 percentage points, achieved completely hands-free.”

Tanya Lenz

“In this experiment environment, it took 19.5 hours to process the 43 parallel trials, running simultaneously across five parallel nodes (allocating one 8x A100 GPU node per search strategy).”

Tanya Lenz

“While a Llama NIM can serve a base model alongside standalone LoRA adapter directories, the Cosmos 3 Reasoner NIM expects a merged checkpoint (base weights fused with the LoRA adapter) rather than a bare adapter.”

Tanya Lenz