Reinforcement learning utilizes reward functions to guide models towards desired behaviors by providing continuous feedback based on defined criteria, a process distinct from supervised learning's reliance on labeled data. This approach allows models to learn through trial and error, as exemplified by reasoning models like DeepSeek-R1. Reward functions, which outline what constitutes a successful outcome, assess model outputs during training and assign scores that inform subsequent iterations, improving performance over time. A tutorial on using reinforcement fine-tuning to train models for the Countdown game highlights the creation of effective reward functions, demonstrating their role in correcting and refining model behavior. The training process on Predibase involves generating model completions, scoring them with reward functions, and feeding ranked outputs back into the loop for enhancement. The use of Chain-of-Thought (CoT) prompting and reward functions like format correctness and proper equation structure significantly improved model accuracy in tasks such as the Countdown game. Reward functions are particularly valuable in scenarios with limited labeled data, such as code generation, strategy games, medical decision support, and personalized AI assistants, and can be dynamically adjusted during training to optimize model performance.