We built a public golf forecasting dataset and trained Golf-Forecaster on resolved tournament questions. The model card reports better Brier score, Brier Skill Score, and calibration than GPT-5 on temporally held-out golf questions.
What we did
- Generated golf forecasting questions from public news context and resolved outcomes.
- Split the data temporally so evaluation used later questions than training.
- Tuned from GPT-OSS-120B with a Brier-score reward and evaluated against GPT-5.
Read more
Primary artifacts for this case study.
Notebook github.com
Golf forecasting fine-tune
Step-by-step notebook for training a golf outcome predictor
Model huggingface.co
Golf-Forecaster
Fine-tuned model for golf tournament outcome prediction
Dataset huggingface.co
Golf forecasting dataset
Tournament questions paired with resolved finishing positions