novastar112/visgym_pacman_2d_random

该数据集包含自定义VisGym Pacman2D环境中随机策略交互的轨迹。每条轨迹包括`image_prev`、`image`和`image_next`的base64 JPEG帧、提示、采样的动作、奖励和环境信息。策略在第一步后以配置的随机停止概率均匀采样移动动作和有效的停止动作。默认生成布局为`data/pacman_2d_ultrahard_v0/train/*.jsonl.gz`，20步 rollout cap，任务配置为`pacman_2d/hard`，`food_count=8`和`max_steps=40`。生成计数、哈希审计和验证元数据位于`metadata/`中。

This dataset contains random-policy interaction trajectories for the custom VisGym Pacman2D environment. Each row is one environment trajectory. The history entries include `image_prev`, `image`, and `image_next` base64 JPEG frames, the prompt shown at the step, the sampled action, reward, and environment info. The policy samples move actions uniformly and samples the valid stop action with the configured random stop probability after the first step. Default generation layout: `data/pacman_2d_ultrahard_v0/train/*.jsonl.gz`, 20-step rollout cap, task config: `pacman_2d/hard` with `food_count=8` and `max_steps=40`. Generation counts, hash audit, and validation metadata are in `metadata/`.