| Field | Response |
|---|---|
| Intended Task/Domain: | Tennis point-level multimodal Q&A, including fine-grained visual recognition (shot mechanics, court positioning, player movement), match-fact and score-state reasoning, rules knowledge, audio cue interpretation, and open-ended point analysis over broadcast and court-capture footage. |
| Model Type: | Transformer |
| Intended Users: | Tennis broadcast and media teams, coaching and performance analysts, sports analytics and production workflows, and enterprise developers building tennis-domain search and assistant applications over archived match video. |
| Output: | Text responses to tennis-domain questions over point-level video inputs (with audio when available), including structured multiple-choice answers and open-ended Q&A; supports chain-of-thought reasoning. |
| Describe how the model works: | The model combines three encoders: a Nemotron Nano V3 LLM backbone (31B MoE, ~3B active), a CRADIO v4-H vision encoder, and a Parakeet speech encoder. Video frames from a point clip are encoded into visual tokens, audio is encoded via the Parakeet encoder when present, and all modality tokens are concatenated with the text question before being processed by the language model to produce a text answer. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable |
| Technical Limitations & Mitigation: | This model supports English only and is optimized for point-level tennis video clips rather than full-match or long-form video. Video input is limited to under 2 minutes at 1 FPS (max 128 frames). Standalone speech transcription, document OCR, and general-purpose agentic workflows are out of scope for this fine-tuned release. Performance may vary on non-tennis content, low-quality or heavily occluded footage, noisy audio, and across different GPU configurations and precision formats. Demographic parity testing across protected classes was not conducted for this fine-tuned tennis release. |
| Verified to have met prescribed NVIDIA quality standards: | Yes |
| Performance Metrics: | Tennis point-level Q&A accuracy (MCQ and open-ended on the Test (unseen matches) partition — fully held-out matches), latency (time-to-first-token (TTFT)), throughput (tokens per second) |
| Potential Known Risks: | The model may generate incorrect or hallucinated answers when processing ambiguous point footage, occluded players, or low-quality video. Performance may degrade on non-tennis content, noisy audio, or inputs that differ from the point-level broadcast and court-capture clips used during fine-tuning. The base model retains pre-training exposure to web-scale text; the model may reflect societal biases or produce unsafe content when prompted adversarially. The model may also generate answers that are inaccurate, omit key information, or include irrelevant text, even when the prompt itself is not explicitly offensive. |
| License/Terms of Use: | Governing Terms: Use of this model is governed by the OpenMDW License Agreement v1.1 (OpenMDW-1.1) |