novastar111/sokoban_medium

Sokoban Medium是一个合成的Sokoban（推箱子）数据集，从本地VisGym Sokoban环境中生成。该数据集主要用于强化学习和规划任务，包含训练和测试轨迹文件，数据规模在10万到100万之间。生成过程中进行了去重处理，确保训练集和测试集之间没有泄漏，最终报告显示数据干净且无重叠。数据集保留了Sokoban环境的视觉资源和动作语义，适用于机器学习和人工智能研究。Sokoban Medium is a synthetic Sokoban dataset generated from the local VisGym Sokoban environment. It is designed for reinforcement learning and planning tasks, containing train and test trajectory files with a size category between 100K and 1M. The generation process includes deduplication to prevent leakage between splits, resulting in a clean dataset with no overlaps. It retains visual assets and action semantics from the integrated Sokoban environment, suitable for machine learning and AI research.