novastar112/sokoban_v0_random

该数据集包含针对自定义VisGym Sokoban环境的随机策略交互轨迹。默认生成布局包括中等任务配置下的20步滚动上限，每个轨迹包含基本64位JPEG帧、提示、采样动作、奖励和环境信息。随机策略均匀采样有效的移动/推动动作，并在第一步之后以配置的随机停止概率采样有效的停止动作。数据集还包含生成计数、哈希审计和验证元数据。

This dataset contains random-policy interaction trajectories for the custom VisGym Sokoban environment. Default generation layout includes a 20-step rollout cap under medium task configuration, with each trajectory containing base64 JPEG frames, prompt, sampled action, reward, and environment info. The random policy samples valid move/push actions uniformly and samples the valid stop action with the configured random stop probability after the first step. The dataset also includes generation counts, hash audit, and validation metadata.