You're setting up RL with verifiable rewards on competition math, starting from a base model prompted to reason in think tags and give its final answer in answer tags.

Before training you run the reward on the base model's zero-shot outputs, and many responses with a plausible answer get reward 0.