Training

Sports intelligence depends on models that can reason in all of the visual, audio, and text modalities. Fine-tuning multimodal language models on domain-specific data is one of the most effective ways to improve accuracy: sports analysis relies on fine-grained details such as player motion, ball trajectory, court geometry, and timing that general-purpose models often miss. Adequate spatial and temporal resolution in both data and model inputs is essential so the model can ground answers in what actually happened in the clip. Getting the most accurate results therefore requires more than tuning the weights of a fixed base model alone. It also depends on how multimodal data is curated, encoded, and aligned across the full fine-tuning pipeline.

This training section covers supervised fine-tuning (SFT) and LoRA (Low-Rank Adaptation) for multimodal language models. The current reference recipes and launch scripts have been tested by the Sports Intelligence Playbooks project with Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16.

Note

Model compatibility: Current playbook recipes target the reference checkpoint above. The NeMo AutoModel stack supports additional models, and support for additional multimodal models is planned as model-specific testing is completed.

See Setup for stack layout and prerequisites. SFT and LoRA link to full runbooks are in the code repo.

• Setup
• SFT
• LoRA