This dataset uses the Gemma-4 thinking format with channel-based reasoning:
<|turn>system
<|think|>You are a helpful assistant...<turn|>
<|turn>user
{problem}<turn|>
<|turn>model
<|channel>thought
{step-by-step reasoning}
<channel|>
{final answer}<turn|>
