NetHack Learning Dataset (NLD)

NetHack Learning Dataset (NLD)是由Meta AI创建的大型数据集，包含从2009年到2020年收集的150万个人类玩家的游戏轨迹和100,000个来自NetHack Challenge 2021的符号机器人获胜者的轨迹。NLD包含三个部分：10亿个状态转换来自人类轨迹，3亿个状态-动作-分数转换来自符号机器人轨迹，以及用于记录、加载和流式传输这些轨迹的配套代码。数据集旨在支持模仿学习、离线强化学习、仅从观察序列学习以及结合离线和在线交互学习的研究。NLD的特点是高效，能够在低内存和计算资源下进行大规模的学习演示。

NetHack Learning Dataset (NLD) is a large-scale dataset created by Meta AI, which contains 1.5 million game trajectories from human players collected between 2009 and 2020, as well as 100,000 winning trajectories of symbolic bots from the 2021 NetHack Challenge. NLD consists of three core components: 1 billion state transitions derived from human player trajectories, 300 million state-action-score transitions from symbolic bot trajectories, and supporting code for recording, loading and streaming these trajectories. This dataset aims to support research on imitation learning, offline reinforcement learning, learning solely from observation sequences, and learning that combines offline and online interactive learning. NLD features high efficiency, enabling large-scale learning demonstrations to be conducted with low memory and computational resources.