WebLINX

该数据集名为WebLINX，包含了指挥官与人类导航员在实时互动中收集的演示数据。它包含了100,000次互动，分布在2,337个由人类专家生成的对话式网页导航演示中。数据集还包括了训练分割，其中有969个演示用于微调，1268个演示用于测试。此外，该数据集还包含了OOD（分布外）样本，用于评估网页导航代理在不同场景下的表现。其规模达到100,000次互动和2,337个演示，任务专注于对话式网页导航。

The dataset is named WebLINX, which encompasses demonstration data collected during real-time interactions between commanders and human navigators. It contains 100,000 interactions across 2,337 conversational web navigation demonstrations generated by human experts. The dataset also includes training splits, with 969 demonstrations used for fine-tuning and 1268 demonstrations allocated for testing. Additionally, the dataset contains OOD (Out-of-Distribution) samples for evaluating the performance of web navigation AI agents across diverse scenarios. With a scale of 100,000 interactions and 2,337 demonstrations, the dataset focuses on conversational web navigation tasks.