ZTE-AIM/Curr-ReFT-data
收藏资源简介：
Curr-ReFT-data数据集包含了为Curr-ReFT方法学两个训练阶段设计的训练数据：Curriculum Reinforcement Learning和Rejected Sample based Self-improvement。数据集分为grpo_data和SFT_data两部分，其中grpo_data包含图像和针对三个任务（视觉检测、视觉分类、多模态数学推理）的训练及测试数据；SFT_data包含用于自我改进的高质量样本。数据集支持在图像和文本上进行问题回答和数学推理任务。
The Curr-ReFT-data dataset contains training data designed for two training stages of the Curr-ReFT methodology: Curriculum Reinforcement Learning and Rejected Sample based Self-improvement. The dataset is divided into grpo_data and SFT_data, where grpo_data includes images and training and testing data for three tasks (visual detection, visual classification, and multimodal mathematical reasoning); SFT_data contains high-quality samples for self-improvement. The dataset supports question answering and mathematical reasoning tasks on images and text.
