IGLU MultiTurn Dataset
收藏arXiv2025-09-30 收录
下载链接：
https://gitlab.aicrowd.com/aicrowd/challenges/iglu-challenge-2022/iglu-2022-rl-mhb-baseline/-/tree/master/nlp_training
下载链接
链接失效反馈官方服务：
资源简介：
该数据集将构建参考结构的复杂IGLU任务分解为多个步骤，使得模型可以在每个步骤上进行训练和评估。数据集的输入包括最新的建筑师指令和语言上下文，而输出则是对应描述放置积木位置的句子。该数据集的规模包括8,736个训练样本，11,283个验证样本以及1,238个测试样本，适用于在类似Minecraft世界的交互式地面语言理解任务。
This dataset decomposes the complex IGLU task of constructing reference structures into multiple sequential steps, enabling models to be trained and evaluated on each individual step. The dataset's inputs include the latest architect instructions and language context, while its outputs are sentences that correspondingly describe the positions of placed building blocks. The dataset contains 8,736 training samples, 11,283 validation samples and 1,238 test samples, and is applicable to interactive grounded language understanding tasks in Minecraft-like worlds.
提供机构：
AICrowd
