Tennis

Tennis is a reference sport for these playbooks and is the primary walk-through example in the sections below.

• Data Curation — annotation, templates, and JSONL splits (avlm/data_prep_example/tennis/)

• Training — SFT/LoRA with the default Nemotron Omni tennis recipes

• Inference — inference on point-level clips

• Evaluation — MCQ and LLM-judge eval (tennis question classes and prompts)

Published checkpoint

Ready-to-run weights are listed on Published Checkpoints — NVIDIA-NemotronLabs-AI-for-Media-Sports-Tennis (in-house manually labeled point clips; ~1.31M Q&A pairs from 43k clips across 239 matches).

Repo entry points

• Data prep README: avlm/data_prep_example/tennis/README.md

• Categorical taxonomy: CATEGORICAL_EVAL.MD

Other sports

Tennis training data in these playbooks is in-house and not redistributed; you can follow the same workflow with your own footage and annotations, and use the published Hugging Face checkpoint for inference and evaluation.

If you want a fully reproducible train-and-eval path on public data, see Basketball. That guide uses the open BARD dataset so you can rerun data preparation, fine-tuning, and scoring end to end with the settings documented there.