R-I-S-E/RISE-Judge-DPO-20K

RISE-Judge-DPO-20k是一个为LLM-as-a-judge设计的偏好数据集，它基于MATH-PRM800K、Ultrafeedback和Skywork-Reward-Preference-80K-v0.2构建而成。该数据集被用于训练judge模型，如R-I-S-E/RISE-Judge-Qwen2.5-32B和R-I-S-E/RISE-Judge-Qwen2.5-7B。数据集的构建是通过在Qwen2.5-32B-Base上提示SFT模型来为MATH-PRM800K、Ultrafeedback和Skywork-Reward-Preference-80K-v0.2中的问题和答案对生成判断，并根据判断的正确性进行样本对的采样。

RISE-Judge-DPO-20k is a preference dataset for LLM-as-a-judge, constructed based on MATH-PRM800K, Ultrafeedback, and Skywork-Reward-Preference-80K-v0.2. It is used to train judge models such as R-I-S-E/RISE-Judge-Qwen2.5-32B and R-I-S-E/RISE-Judge-Qwen2.5-7B. The dataset is curated by prompting an SFT model on Qwen2.5-32B-Base to generate judgments for question-answer pairs in MATH-PRM800K, Ultrafeedback, and Skywork-Reward-Preference-80K-v0.2, and sampling answer pairs according to the correctness of the judgments.