training data json file release

[justgogorunrun]

Hello, I noticed that the article shows the training dataset format as“ {a_1} think /think {a_2}”, but the currently released JSON on HuggingFace only contains the answer without the thinking part. Has the full JSON file not been released yet?

[sming256]

Thanks for the question.

We skip the SFT stage and train with RL directly. We don’t have annotated thinking traces, so the released JSON includes answers only, and our RL uses verifiable rewards computed on the two answers.