SI-Lab/StepCountQA-RL-Traj_11_50_Combined

StepCountQA-RL-Traj_11_50_Combined是一个统一的视觉计数RL训练数据集，覆盖了从11到50的对象计数。它合并了两个互补的11-50数据集：StepCountQA-RL-Traj_11_50（9,508个样本，带有JSON注释轨迹，转换为数字）和StepCountQA-RL-Traj_11_50_NumericOnly（29,722个样本，纯数字答案来自SFT轨迹）。合并后的数据集共有39,230个样本，无重复。每个样本包含一个原始图像、一个计数问题和一个纯数字答案字符串。数据集的结构包括图像序列、问题字符串和答案字符串。数据集的统计信息显示，总样本数为39,230，计数范围为11到50，无重复图像，总大小约为8.95 GB。计数分布表详细列出了每个计数对应的样本数量。数据集的使用方法是通过HuggingFace的load_dataset函数加载。数据集的创建过程包括从两个源数据集提取图像路径字段，验证无重复，然后合并为4个输出分片。

StepCountQA-RL-Traj_11_50_Combined is a unified visual counting RL training dataset covering object counts from 11 to 50. It merges two complementary 11-50 datasets: StepCountQA-RL-Traj_11_50 (9,508 samples with `<point>` JSON annotation trajectories, converted to numeric) and StepCountQA-RL-Traj_11_50_NumericOnly (29,722 samples with pure numeric answers from SFT trajectories). The combined dataset contains 39,230 samples with zero overlap. Each sample pairs an original image with a counting question and a pure numeric answer string. The dataset structure includes a sequence of images, a problem string, and an answer string. Statistics show 39,230 total samples, count range 11–50, zero duplicate images, and ~8.95 GB total size. A count distribution table details sample counts per number. Usage involves loading via HuggingFaces `load_dataset`. Creation involved merging two sources with verified zero overlap into 4 output shards.