brahmandam/DoomFrameDataset

DoomFrameDataset是一个基于ViZDoom游戏环境生成的帧-动作数据集，用于强化学习、模仿学习、行为克隆和离线强化学习实验。数据集以WebDataset格式打包，包含RGB游戏帧与策略选择的动作配对，以及每步的元数据（如奖励、回合ID、步骤ID、终止标志和价值估计）。数据集分为preview（预览版，约79k样本，用于快速检查）和full（完整版，约240万样本，用于训练）两个配置，总大小约68 GB。数据文件包括多个tar分片和action_map.json，每个分片包含配对的PNG图像和JSON元数据文件。该数据集通过自动化ViZDoom策略滚动生成，应视为游戏观察/动作数据，而非人类演示。

DoomFrameDataset is a ViZDoom frame-action dataset generated from policy rollouts, packaged as WebDataset tar shards for streaming training, imitation learning, behavior cloning, and offline reinforcement-learning experiments. The dataset contains RGB game frames paired with the action selected by the rollout policy and per-step metadata such as reward, episode id, step id, terminal flag, and value estimate. It includes two configurations: preview (with ~79k samples for quick checks) and full (with 2,398,745 samples for training), totaling about 68 GB. The files consist of multiple tar shards and action_map.json, each shard containing paired PNG images and JSON metadata files. The dataset was generated from automated ViZDoom policy rollouts and should be treated as gameplay observation/action data, not human demonstrations.