CONAN

该数据集名为Conan，是一个专为评估主动推理能力而设计的互动开放世界环境。它不仅促进了积极的探索行为，还推动了多轮演绎推理的发展。在训练过程中，研究者采用了多种强化学习算法来训练探险者，并根据探险者的表现，对不同的推理模型设置进行了评估。该数据集的训练规模宏大，包含了大量的探索步骤（达到了10的8次方）。其任务是在开放世界环境中进行主动推理。

This dataset, named Conan, is an interactive open-world environment specifically designed for evaluating active reasoning capabilities. It not only fosters proactive exploratory behaviors but also advances the development of multi-turn deductive reasoning. During the training process, researchers adopted multiple reinforcement learning algorithms to train the explorers, and evaluated different reasoning model configurations based on the explorers' performance. The training scale of this dataset is massive, containing a vast number of exploration steps (up to 10^8). Its core task is to conduct active reasoning within open-world environments.