LMRL-Gym

LMRL-Gym数据集由加州大学伯克利分校创建，旨在为多轮强化学习提供基准测试。该数据集包含8个语言任务，涉及开放式对话和文本游戏，要求多个轮次的语言交互。数据集通过大型语言模型生成的合成数据创建，支持离线强化学习训练，并提供模拟器用于评估训练后的代理在多轮交互任务中的性能。数据集的应用领域包括复杂对话、游戏和工具使用，旨在解决语言模型在目标导向推理和规划中的挑战。The LMRL-Gym dataset, developed by the University of California, Berkeley, serves as a benchmark for multi-turn reinforcement learning. It comprises 8 language tasks spanning open-domain dialogue and text-based games, all requiring multi-turn linguistic interactions. Constructed from synthetic data generated by large language models, this dataset supports offline reinforcement learning training and offers simulators to evaluate the performance of trained AI agents in multi-turn interactive tasks. Its application areas include complex dialogue, gaming, and tool use, with the goal of addressing the challenges that language models encounter in goal-oriented reasoning and planning.