I have an instance of NVIDIA’s COSMOS-Reason1-7B that already ingests a short video clip and returns a JSON payload. I now need it to excel at one very specific domain: soccer. Over the next 6–8 hours I want to: • scrape or otherwise source open-source sports footage (soccer only), • label the clips for Penalty Shot, Goal, Red Card, Yellow Card, Hat-Trick and any other easily identifiable key moments that fit the same pattern, • run a rapid fine-tune on COSMOS-Reason1-7B so that the model reliably emits a clean JSON object describing those events, and • hand back the fine-tuned weights, the annotation set, the training script (PyTorch + CUDA is fine) and a quick demo notebook or script that shows the before/after performance on two test clips. The model is already set up on an NVIDIA GPU box with all dependencies; I simply need your data-wrangling, annotation and tuning skills to deliver the improved checkpoint and proof it works. If you have a preferred framework such as LoRA or QLoRA for fast tuning, feel free to use it—speed is key today.