A team is developing a chatbot to provide general health information. A critical safety requirement is that the chatbot must never suggest specific medication dosages. However, during testing, the model sometimes generates outputs like, 'For a headache, you could take 500mg of paracetamol.' Your task is to use a penalty function during the decoding process to prevent this behavior. Describe the specific characteristics or patterns in the generated text that your penalty function would be designed to identify and penalize to solve this problem.